torch-rb 0.9.2 → 0.10.0

data/codegen/native_functions.yaml

@@ -100,10 +100,49 @@
 
 - func: _make_dual(Tensor(a) primal, Tensor tangent, int level) -> Tensor(a)
   variants: function
+  dispatch:
+    CompositeExplicitAutograd: _make_dual
 
 - func: _unpack_dual(Tensor(a) dual, int level) -> (Tensor(a) primal, Tensor tangent)
   variants: function
 
+# NOTE: [_new_zeros_with_same_feature_meta]
+# This function creates a new tensor with the layout and TensorOptions
+# of `other` but also takes into account the batch dimensions of `self`
+#
+# This function has a couple extra constraints because it is also used for `jvp`
+# in functorch.
+# - is used for forward AD because there is the restriction
+#   that the primal and tangent must have the same layout
+# - We cannot assume that `self` and `other` have the same sizes or even dim
+#   because in the inplace over view case, `other` is the base tensor, and
+#   `self` is the forward grad with respect to the view, which can have an
+#   entirely different shape
+# - takes the number of batch dims for `self` because we also handle
+#   some batching logic. We handle that here instead of a batching rule because
+#   we'd like to avoid calling as_strided in the batching rule (as to enable
+#   nested vmap in functorch).
+# - needs to be CompositeExplicitAutograd for jvp support in functorch.
+#   functorch currently relies on TensorWrapper which does not have storage
+#   CompositeExplicitAutograd makes sure the TensorWrapper is unwrapped.
+# - this function may eventually take on another int argument to store the
+#   the number of batch dims for other once we support that use case
+- func: _new_zeros_with_same_feature_meta(Tensor self, Tensor other, *, int self_num_batch_dims=0) -> Tensor
+  variants: function
+  dispatch:
+    CompositeExplicitAutograd: _new_zeros_with_same_feature_meta
+
+# This function compares the storage numel of self with that of other, where
+# storage numel is cumputed as: `other.storage().nbytes() / other.itemsize()`.
+# We create this function for composite compliance purposes. The batching rule
+# always returns true because vmapped as_strided does not support accessing
+# storage locations not indexable by the input tensor.
+# See the note above for more information.
+- func: _has_same_storage_numel(Tensor self, Tensor other) -> bool
+  variants: function
+  dispatch:
+    CompositeExplicitAutograd: _has_same_storage_numel
+
 - func: rename_(Tensor(a!) self, Dimname[]? names) -> Tensor(a!)
   variants: method
 
@@ -176,6 +215,17 @@
   dispatch:
     CUDA: masked_scale_cuda
 
+- func: native_dropout(Tensor input, float p, bool? train) -> (Tensor, Tensor)
+  variants: function
+  dispatch:
+    CPU: native_dropout_cpu
+    CUDA: native_dropout_cuda
+
+- func: native_dropout_backward(Tensor grad_output, Tensor mask, float scale) -> Tensor
+  dispatch:
+    CPU: native_dropout_backward_cpu
+    CUDA: native_dropout_backward_cuda
+
 - func: _sobol_engine_draw(Tensor quasi, int n, Tensor sobolstate, int dimension, int num_generated, ScalarType? dtype) -> (Tensor, Tensor)
 
 - func: _sobol_engine_ff_(Tensor(a!) self, int n, Tensor sobolstate, int dimension, int num_generated) -> Tensor(a!)
@@ -209,17 +259,23 @@
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: abs
+    SparseCPU, SparseCUDA: abs_sparse
+    SparseCsrCPU, SparseCsrCUDA: abs_sparse_csr
 
 - func: abs_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: abs_
+    SparseCPU, SparseCUDA: abs_sparse_
+    SparseCsrCPU, SparseCsrCUDA: abs_sparse_csr_
 
 - func: abs.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   dispatch:
     CPU, CUDA: abs_out
+    SparseCPU, SparseCUDA: abs_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: abs_sparse_csr_out
 
 # Note [Adding an alias]
 # To add an alias do the following:
@@ -231,18 +287,15 @@
 #      will stop it from "inheriting" the original operation's autograd behavior.
 # 2) Implement the corresponding functions and have them redispatch to the
 #      original function.
-# 3) Add entries for the alias (and original function, if needed) to
-#      aten/src/ATen/core/interned_strings.h
-#      (This may require removing an entry from ATen/core/aten_interned_strings.h.)
-# 4) Add docstrings to the new function that reference the original function,
+# 3) Add docstrings to the new function that reference the original function,
 #      and document the method as usual (if it exists.)
 #    (See torch/_torch_docs.py and docs/source/torch.rst if adding a function,
 #     torch/_tensor_docs.py and docs/source/tensors.rst if adding a method,
 #     or module-specific doc bindings (like torch/linalg/__init__.py) if
 #     adding an alias in a namespace.)
-# 5) Update torch/overrides.py consistent with the original function.
-# 6) Update the alias_map in torch/csrc/jit/passes/normalize_ops.cpp.
-# 7) Add aliases argument to existing OpInfo/UnaryUfuncInfo or create new OpInfo/UnaryUfuncInfo entry
+# 4) Update torch/overrides.py consistent with the original function.
+# 5) Update the alias_map in torch/csrc/jit/passes/normalize_ops.cpp.
+# 6) Add aliases argument to existing OpInfo/UnaryUfuncInfo or create new OpInfo/UnaryUfuncInfo entry
 # in op_db list in torch/testing/_internal/common_methods_invocations.py
 #
 # See torch.absolute, an alias for torch.abs, as an example.
@@ -264,11 +317,13 @@
   variants: function, method
   dispatch:
     CPU, CUDA: angle
+    SparseCsrCPU, SparseCsrCUDA: angle_sparse_csr
 
 - func: angle.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   dispatch:
     CPU, CUDA: angle_out
+    SparseCsrCPU, SparseCsrCUDA: angle_sparse_csr_out
 
 - func: view_as_real(Tensor(a) self) -> Tensor(a)
   variants: function
@@ -283,16 +338,24 @@
 - func: sgn(Tensor self) -> Tensor
   variants: function, method
   structured_delegate: sgn.out
+  dispatch:
+    SparseCPU, SparseCUDA: sgn_sparse
+    SparseCsrCPU, SparseCsrCUDA: sgn_sparse_csr
 
 - func: sgn_(Tensor(a!) self) -> Tensor(a!)
   variants: method
   structured_delegate: sgn.out
+  dispatch:
+    SparseCPU, SparseCUDA: sgn_sparse_
+    SparseCsrCPU, SparseCsrCUDA: sgn_sparse_csr_
 
 - func: sgn.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: sgn_out
+    SparseCPU, SparseCUDA: sgn_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: sgn_sparse_csr_out
 
 - func: real(Tensor(a) self) -> Tensor(a)
   device_check: NoCheck   # TensorIterator
@@ -315,6 +378,7 @@
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: _conj_physical
+    SparseCsrCPU, SparseCsrCUDA: conj_physical_sparse_csr
 
 - func: conj_physical(Tensor self) -> Tensor
   variants: function, method
@@ -323,11 +387,13 @@
   dispatch:
     CPU, CUDA: conj_physical_out
     SparseCPU, SparseCUDA: conj_physical_out_sparse
+    SparseCsrCPU, SparseCsrCUDA: conj_physical_sparse_csr_out
 
 - func: conj_physical_(Tensor(a!) self) -> Tensor(a!)
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: conj_physical_
+    SparseCsrCPU, SparseCsrCUDA: conj_physical_sparse_csr_
 
 - func: resolve_conj(Tensor(a) self) -> Tensor(a)
   variants: function, method
@@ -381,6 +447,7 @@
     SparseCPU, SparseCUDA: add_sparse
     SparseCsrCPU, SparseCsrCUDA: add_sparse_csr
     MkldnnCPU: mkldnn_add
+    ZeroTensor: add_zerotensor
 
 - func: add_.Tensor(Tensor(a!) self, Tensor other, *, Scalar alpha=1) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -454,6 +521,8 @@
   dispatch:
     CPU: addmv_out_cpu
     CUDA: addmv_out_cuda
+    SparseCsrCPU: addmv_out_sparse_csr
+    SparseCsrCUDA: addmv_out_sparse_csr_cuda
 
 - func: addr(Tensor self, Tensor vec1, Tensor vec2, *, Scalar beta=1, Scalar alpha=1) -> Tensor
   variants: function, method
@@ -532,7 +601,7 @@
 
 - func: arange.start_out(Scalar start, Scalar end, Scalar step=1, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
-    CPU: arange_cpu_out
+    CPU, Meta: arange_out
     CUDA: arange_cuda_out
 
 # This function is a temporary hack to allow tracing of arange like constructs with dynamic
@@ -588,16 +657,24 @@
 - func: asinh(Tensor self) -> Tensor
   variants: function, method
   structured_delegate: asinh.out
+  dispatch:
+    SparseCPU, SparseCUDA: asinh_sparse
+    SparseCsrCPU, SparseCsrCUDA: asinh_sparse_csr
 
 - func: asinh_(Tensor(a!) self) -> Tensor(a!)
   variants: function, method
   structured_delegate: asinh.out
+  dispatch:
+    SparseCPU, SparseCUDA: asinh_sparse_
+    SparseCsrCPU, SparseCsrCUDA: asinh_sparse_csr_
 
 - func: asinh.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: asinh_out
+    SparseCPU, SparseCUDA: asinh_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: asinh_sparse_csr_out
 
 # arcsinh, alias for asinh
 - func: arcsinh(Tensor self) -> Tensor
@@ -611,16 +688,25 @@
 - func: atanh(Tensor self) -> Tensor
   structured_delegate: atanh.out
   variants: function, method
+  dispatch:
+    CompositeExplicitAutograd: atanh
+    SparseCPU, SparseCUDA: atanh_sparse
+    SparseCsrCPU, SparseCsrCUDA: atanh_sparse_csr
 
 - func: atanh_(Tensor(a!) self) -> Tensor(a!)
   structured_delegate: atanh.out
   variants: function, method
+  dispatch:
+    SparseCPU, SparseCUDA: atanh_sparse_
+    SparseCsrCPU, SparseCsrCUDA: atanh_sparse_csr_
 
 - func: atanh.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: atanh_out
+    SparseCPU, SparseCUDA: atanh_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: atanh_sparse_csr_out
 
 # arctanh, alias for atanh
 - func: arctanh(Tensor self) -> Tensor
@@ -634,7 +720,7 @@
 - func: as_strided(Tensor(a) self, int[] size, int[] stride, int? storage_offset=None) -> Tensor(a)
   variants: function, method
   dispatch:
-    CPU, CUDA, Meta: as_strided_tensorimpl
+    ZeroTensor, CPU, CUDA, Meta: as_strided_tensorimpl
     QuantizedCPU, QuantizedCUDA: as_strided_qtensorimpl
   device_check: NoCheck
   device_guard: False
@@ -644,6 +730,7 @@
   variants: function, method
   device_check: NoCheck
   device_guard: False
+  tags: inplace_view
   dispatch:
     CompositeExplicitAutograd: as_strided_
 
@@ -653,6 +740,7 @@
   structured_delegate: asin.out
   dispatch:
     SparseCPU, SparseCUDA: asin_sparse
+    SparseCsrCPU, SparseCsrCUDA: asin_sparse_csr
 
 - func: asin_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -660,6 +748,7 @@
   structured_delegate: asin.out
   dispatch:
     SparseCPU, SparseCUDA: asin_sparse_
+    SparseCsrCPU, SparseCsrCUDA: asin_sparse_csr_
 
 - func: asin.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -667,7 +756,8 @@
   structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: asin_out
-    SparseCPU, SparseCUDA: asin_out_sparse
+    SparseCPU, SparseCUDA: asin_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: asin_sparse_csr_out
 
 # arcsin, alias of asin
 - func: arcsin(Tensor self) -> Tensor
@@ -682,11 +772,17 @@
   device_check: NoCheck   # TensorIterator
   structured_delegate: atan.out
   variants: function, method
+  dispatch:
+    SparseCPU, SparseCUDA: atan_sparse
+    SparseCsrCPU, SparseCsrCUDA: atan_sparse_csr
 
 - func: atan_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   structured_delegate: atan.out
   variants: function, method
+  dispatch:
+    SparseCPU, SparseCUDA: atan_sparse_
+    SparseCsrCPU, SparseCsrCUDA: atan_sparse_csr_
 
 - func: atan.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -694,6 +790,8 @@
   structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: atan_out
+    SparseCPU, SparseCUDA: atan_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: atan_sparse_csr_out
 
 # arctan, alias of atan
 - func: arctan(Tensor self) -> Tensor
@@ -723,24 +821,19 @@
 
 - func: baddbmm(Tensor self, Tensor batch1, Tensor batch2, *, Scalar beta=1, Scalar alpha=1) -> Tensor
   variants: function, method
-  dispatch:
-    CPU: baddbmm_cpu
-    CUDA: baddbmm_cuda
+  structured_delegate: baddbmm.out
 
 - func: baddbmm_(Tensor(a!) self, Tensor batch1, Tensor batch2, *, Scalar beta=1, Scalar alpha=1) -> Tensor(a!)
   variants: method
-  dispatch:
-    CPU: baddbmm__cpu
-    CUDA: baddbmm__cuda
-
-- func: _baddbmm_mkl_(Tensor(a!) self, Tensor batch1, Tensor batch2, *, Scalar beta=1, Scalar alpha=1) -> Tensor(a!)
-  variants: function
+  structured_delegate: baddbmm.out
 
 - func: baddbmm.out(Tensor self, Tensor batch1, Tensor batch2, *, Scalar beta=1, Scalar alpha=1, Tensor(a!) out) -> Tensor(a!)
+  structured: True
   variants: function
   dispatch:
     CPU: baddbmm_out_cpu
     CUDA: baddbmm_out_cuda
+    SparseCsrCUDA: baddbmm_out_sparse_csr_cuda
 
 - func: bartlett_window(int window_length, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
 
@@ -788,7 +881,7 @@
   device_check: NoCheck   # TensorIterator
   variants: function, method
 
-- func: bilinear(Tensor input1, Tensor input2, Tensor weight, Tensor? bias) -> Tensor
+- func: bilinear(Tensor input1, Tensor input2, Tensor weight, Tensor? bias=None) -> Tensor
 
 - func: binary_cross_entropy(Tensor self, Tensor target, Tensor? weight=None, int reduction=Mean) -> Tensor
   device_check: NoCheck   # TensorIterator
@@ -886,10 +979,14 @@
 - func: logical_not(Tensor self) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
+  dispatch:
+    CompositeExplicitAutograd: logical_not
 
 - func: logical_not_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   variants: method
+  dispatch:
+    CompositeExplicitAutograd: logical_not_
 
 - func: logical_not.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -899,10 +996,14 @@
 - func: logical_xor(Tensor self, Tensor other) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
+  dispatch:
+    CompositeExplicitAutograd: logical_xor
 
 - func: logical_xor_(Tensor(a!) self, Tensor other) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   variants: method
+  dispatch:
+    CompositeExplicitAutograd: logical_xor_
 
 - func: logical_xor.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -912,10 +1013,14 @@
 - func: logical_and(Tensor self, Tensor other) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
+  dispatch:
+    CompositeExplicitAutograd: logical_and
 
 - func: logical_and_(Tensor(a!) self, Tensor other) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   variants: method
+  dispatch:
+    CompositeExplicitAutograd: logical_and_
 
 - func: logical_and.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -925,10 +1030,14 @@
 - func: logical_or(Tensor self, Tensor other) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
+  dispatch:
+    CompositeExplicitAutograd: logical_or
 
 - func: logical_or_(Tensor(a!) self, Tensor other) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   variants: method
+  dispatch:
+    CompositeExplicitAutograd: logical_or_
 
 - func: logical_or.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -940,20 +1049,21 @@
 - func: blackman_window.periodic(int window_length, bool periodic, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
 
 - func: bmm(Tensor self, Tensor mat2) -> Tensor
+  structured_delegate: bmm.out
   variants: function, method
   dispatch:
-    CPU: bmm_cpu
-    CUDA: bmm_cuda
     SparseCPU: bmm_sparse_cpu
     SparseCUDA: bmm_sparse_cuda
 
 - func: bmm.out(Tensor self, Tensor mat2, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
   variants: function
   dispatch:
     CPU: bmm_out_cpu
     CUDA: bmm_out_cuda
     SparseCPU: bmm_out_sparse_cpu
     SparseCUDA: bmm_out_sparse_cuda
+    SparseCsrCUDA: bmm_out_sparse_csr_cuda
 
 - func: broadcast_tensors(Tensor[] tensors) -> Tensor[]
   device_check: NoCheck
@@ -962,6 +1072,11 @@
 - func: broadcast_to(Tensor(a) self, int[] size) -> Tensor(a)
   variants: function, method
 
+- func: _sparse_broadcast_to(Tensor(a) self, int[] size) -> Tensor(a)
+  variants: function
+  dispatch:
+    SparseCPU, SparseCUDA: sparse_broadcast_to
+
 - func: cat(Tensor[] tensors, int dim=0) -> Tensor
   dispatch:
     CompositeExplicitAutograd: cat
@@ -992,6 +1107,8 @@
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: ceil
+    SparseCPU, SparseCUDA: ceil_sparse
+    SparseCsrCPU, SparseCsrCUDA: ceil_sparse_csr
 
 - func: ceil_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -999,6 +1116,8 @@
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: ceil_
+    SparseCPU, SparseCUDA: ceil_sparse_
+    SparseCsrCPU, SparseCsrCUDA: ceil_sparse_csr_
 
 - func: ceil.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -1006,6 +1125,8 @@
   structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: ceil_out
+    SparseCPU, SparseCUDA: ceil_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: ceil_sparse_csr_out
 
 # alias for torch.linalg.multi_dot
 - func: chain_matmul(Tensor[] matrices) -> Tensor
@@ -1019,18 +1140,18 @@
   device_check: NoCheck
   device_guard: False
 
-- func: chunk(Tensor(a) self, int chunks, int dim=0) -> Tensor(a)[]
+- func: chunk(Tensor(a -> *) self, int chunks, int dim=0) -> Tensor(a)[]
   variants: function, method
   device_check: NoCheck
   device_guard: False
 
-- func: tensor_split.sections(Tensor(a) self, int sections, int dim=0) -> Tensor(a)[]
+- func: tensor_split.sections(Tensor(a -> *) self, int sections, int dim=0) -> Tensor(a)[]
   variants: function, method
 
-- func: tensor_split.indices(Tensor(a) self, int[] indices, int dim=0) -> Tensor(a)[]
+- func: tensor_split.indices(Tensor(a -> *) self, int[] indices, int dim=0) -> Tensor(a)[]
   variants: function, method
 
-- func: tensor_split.tensor_indices_or_sections(Tensor(a) self, Tensor tensor_indices_or_sections, int dim=0) -> Tensor(a)[]
+- func: tensor_split.tensor_indices_or_sections(Tensor(a -> *) self, Tensor tensor_indices_or_sections, int dim=0) -> Tensor(a)[]
   variants: function, method
 
 - func: clamp(Tensor self, Scalar? min=None, Scalar? max=None) -> Tensor
@@ -1186,6 +1307,12 @@
   manual_cpp_binding: True
 
 - func: convolution(Tensor input, Tensor weight, Tensor? bias, int[] stride, int[] padding, int[] dilation, bool transposed, int[] output_padding, int groups) -> Tensor
+  dispatch:
+    CompositeExplicitAutograd: convolution
+
+- func: convolution_backward(Tensor grad_output, Tensor input, Tensor weight, int[]? bias_sizes, int[] stride, int[] padding, int[] dilation, bool transposed, int[] output_padding, int groups, bool[3] output_mask) -> (Tensor, Tensor, Tensor)
+  dispatch:
+    CompositeExplicitAutograd, CUDA: convolution_backward
 
 - func: convolution_overrideable(Tensor input, Tensor weight, Tensor? bias, int[] stride, int[] padding, int[] dilation, bool transposed, int[] output_padding, int groups) -> Tensor
   dispatch:
@@ -1196,14 +1323,14 @@
     CompositeExplicitAutograd: convolution_backward_overrideable
 
 - func: _convolution(Tensor input, Tensor weight, Tensor? bias, int[] stride, int[] padding, int[] dilation, bool transposed, int[] output_padding, int groups, bool benchmark, bool deterministic, bool cudnn_enabled, bool allow_tf32) -> Tensor
+  dispatch:
+    CompositeExplicitAutograd: _convolution
 
 - func: _convolution.deprecated(Tensor input, Tensor weight, Tensor? bias, int[] stride, int[] padding, int[] dilation, bool transposed, int[] output_padding, int groups, bool benchmark, bool deterministic, bool cudnn_enabled) -> Tensor
 
 - func: _convolution_mode(Tensor input, Tensor weight, Tensor? bias, int[] stride, str padding, int[] dilation, int groups) -> Tensor
 
-- func: _convolution_nogroup(Tensor input, Tensor weight, Tensor? bias, int[] stride, int[] padding, int[] dilation, bool transposed, int[] output_padding) -> Tensor
-
-- func: _convolution_double_backward(Tensor? ggI, Tensor? ggW, Tensor? ggb, Tensor gO, Tensor weight, Tensor self, int[] stride, int[] padding, int[] dilation, bool transposed, int[] output_padding, int groups, bool benchmark, bool deterministic, bool cudnn_enabled, bool allow_tf32, bool[3] output_mask) -> (Tensor, Tensor, Tensor)
+- func: _convolution_double_backward(Tensor? ggI, Tensor? ggW, Tensor? ggb, Tensor gO, Tensor weight, Tensor self, int[] stride, int[] padding, int[] dilation, bool transposed, int[] output_padding, int groups, bool[3] output_mask) -> (Tensor, Tensor, Tensor)
 
 - func: conv1d(Tensor input, Tensor weight, Tensor? bias=None, int[1] stride=1, int[1] padding=0, int[1] dilation=1, int groups=1) -> Tensor
 
@@ -1239,7 +1366,9 @@
   device_guard: False
   dispatch:
     MkldnnCPU: copy_mkldnn_
+    SparseCPU, SparseCUDA, SparseHIP: copy_sparse_wrapper_
     CompositeExplicitAutograd: copy_
+    SparseCsrCPU, SparseCsrCUDA: copy_sparse_csr_
 
 - func: _copy_from(Tensor self, Tensor dst, bool non_blocking=False) -> Tensor
   dispatch: {}
@@ -1320,56 +1449,14 @@
   dispatch:
     CUDA: cudnn_batch_norm_backward
 
-- func: cudnn_convolution.deprecated(Tensor self, Tensor weight, Tensor? bias, int[] padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic) -> Tensor
-  dispatch:
-    CUDA: cudnn_convolution_deprecated
-
-- func: cudnn_convolution.deprecated2(Tensor self, Tensor weight, int[] padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic) -> Tensor
-  dispatch:
-    CUDA: cudnn_convolution_deprecated2
-
 - func: cudnn_convolution(Tensor self, Tensor weight, int[] padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic, bool allow_tf32) -> Tensor
   dispatch:
     CUDA: cudnn_convolution
 
-- func: cudnn_convolution_backward_input(int[] self_size, Tensor grad_output, Tensor weight, int[] padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic, bool allow_tf32) -> Tensor
-  dispatch:
-    CUDA: cudnn_convolution_backward_input
-
-- func: cudnn_convolution_backward(Tensor self, Tensor grad_output, Tensor weight, int[] padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic, bool allow_tf32, bool[2] output_mask) -> (Tensor, Tensor)
-  dispatch:
-    CUDA: cudnn_convolution_backward
-
-- func: cudnn_convolution_backward_weight(int[] weight_size, Tensor grad_output, Tensor self, int[] padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic, bool allow_tf32) -> Tensor
-  dispatch:
-    CUDA: cudnn_convolution_backward_weight
-
-- func: cudnn_convolution_transpose.deprecated(Tensor self, Tensor weight, Tensor? bias, int[] padding, int[] output_padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic) -> Tensor
-  dispatch:
-    CUDA: cudnn_convolution_transpose_deprecated
-
-- func: cudnn_convolution_transpose.deprecated2(Tensor self, Tensor weight, int[] padding, int[] output_padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic) -> Tensor
-  dispatch:
-    CUDA: cudnn_convolution_transpose_deprecated2
-
 - func: cudnn_convolution_transpose(Tensor self, Tensor weight, int[] padding, int[] output_padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic, bool allow_tf32) -> Tensor
   dispatch:
     CUDA: cudnn_convolution_transpose
 
-# NB: output_padding not strictly needed here, but it's helpful for the float
-# backwards
-- func: cudnn_convolution_transpose_backward(Tensor self, Tensor grad_output, Tensor weight, int[] padding, int[] output_padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic, bool allow_tf32, bool[2] output_mask) -> (Tensor, Tensor)
-  dispatch:
-    CUDA: cudnn_convolution_transpose_backward
-
-- func: cudnn_convolution_transpose_backward_input(Tensor grad_output, Tensor weight, int[] padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic, bool allow_tf32) -> Tensor
-  dispatch:
-    CUDA: cudnn_convolution_transpose_backward_input
-
-- func: cudnn_convolution_transpose_backward_weight(int[] weight_size, Tensor grad_output, Tensor self, int[] padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic, bool allow_tf32) -> Tensor
-  dispatch:
-    CUDA: cudnn_convolution_transpose_backward_weight
-
 - func: cudnn_convolution_relu(Tensor self, Tensor weight, Tensor? bias, int[] stride, int[] padding, int[] dilation, int groups) -> Tensor
   dispatch:
     CUDA: cudnn_convolution_relu
@@ -1516,6 +1603,8 @@
 
 - func: diag_embed(Tensor self, int offset=0, int dim1=-2, int dim2=-1) -> Tensor
   variants: function, method
+  dispatch:
+    CompositeExplicitAutograd: diag_embed
 
 - func: diagflat(Tensor self, int offset=0) -> Tensor
   variants: function, method
@@ -1525,6 +1614,10 @@
   dispatch:
     CompositeExplicitAutograd: diagonal
 
+- func: linalg_diagonal(Tensor(a) A, *, int offset=0, int dim1=-2, int dim2=-1) -> Tensor(a)
+  python_module: linalg
+  variants: function
+
 - func: diagonal.Dimname(Tensor(a) self, *, Dimname outdim, Dimname dim1, Dimname dim2, int offset=0) -> Tensor(a)
   variants: function, method
 
@@ -1571,6 +1664,7 @@
   structured_delegate: div.out
   dispatch:
     SparseCPU, SparseCUDA: div_sparse
+    ZeroTensor: div_zerotensor
 
 - func: div_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -1781,12 +1875,19 @@
     Meta: empty_meta
     MkldnnCPU: empty_mkldnn
     SparseCPU, SparseCUDA: empty_sparse
+    SparseCsrCPU, SparseCsrCUDA: empty_sparse_csr
 
+# We do not make new_empty a composite that calls into new_empty_strided, as the strided version
+# is significantly more difficult to implement by different backends
 - func: new_empty(Tensor self, int[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
   variants: method
+  dispatch:
+    CompositeExplicitAutograd: new_empty
 
 - func: new_empty_strided(Tensor self, int[] size, int[] stride, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
   variants: method
+  dispatch:
+    CompositeExplicitAutograd: new_empty_strided
 
 - func: new_full(Tensor self, int[] size, Scalar fill_value, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
   variants: method
@@ -1820,6 +1921,7 @@
     CPU, Meta: resize_
     CUDA: resize_cuda_
     QuantizedCPU: quantized_resize_cpu_
+    SparseCsrCPU, SparseCsrCUDA: resize_sparse_csr_
 
 - func: empty_quantized(int[] size, Tensor qtensor, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None, MemoryFormat? memory_format=None) -> Tensor
   category_override: factory
@@ -1834,6 +1936,10 @@
 - func: empty_like(Tensor self, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None, MemoryFormat? memory_format=None) -> Tensor
   device_check: NoCheck
   device_guard: False
+  dispatch:
+    CompositeExplicitAutograd: empty_like
+    SparseCPU, SparseCUDA: empty_like_sparse_coo
+    SparseCsrCPU, SparseCsrCUDA: empty_like_sparse_csr
 
 - func: empty_strided(int[] size, int[] stride, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
   dispatch:
@@ -1845,11 +1951,17 @@
   device_check: NoCheck   # TensorIterator
   structured_delegate: erf.out
   variants: function, method
+  dispatch:
+    SparseCPU, SparseCUDA: erf_sparse
+    SparseCsrCPU, SparseCsrCUDA: erf_sparse_csr
 
 - func: erf_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   structured_delegate: erf.out
   variants: function, method
+  dispatch:
+    SparseCPU, SparseCUDA: erf_sparse_
+    SparseCsrCPU, SparseCsrCUDA: erf_sparse_csr_
 
 - func: erf.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -1857,6 +1969,8 @@
   structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: erf_out
+    SparseCPU, SparseCUDA: erf_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: erf_sparse_csr_out
 
 - func: erfc(Tensor self) -> Tensor
   device_check: NoCheck   # TensorIterator
@@ -1910,11 +2024,17 @@
   device_check: NoCheck   # TensorIterator
   structured_delegate: expm1.out
   variants: function, method
+  dispatch:
+    SparseCPU, SparseCUDA: expm1_sparse
+    SparseCsrCPU, SparseCsrCUDA: expm1_sparse_csr
 
 - func: expm1_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   structured_delegate: expm1.out
   variants: function, method
+  dispatch:
+    SparseCPU, SparseCUDA: expm1_sparse_
+    SparseCsrCPU, SparseCsrCUDA: expm1_sparse_csr_
 
 - func: expm1.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -1922,6 +2042,8 @@
   structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: expm1_out
+    SparseCPU, SparseCUDA: expm1_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: expm1_sparse_csr_out
 
 - func: expand(Tensor(a) self, int[] size, *, bool implicit=False) -> Tensor(a)
   variants: method  # This is method-only to match the previous tensor API. In the future we could make this a function too.
@@ -1971,14 +2093,16 @@
   device_check: NoCheck   # TensorIterator
   variants: function, method
   dispatch:
-    CPU, CUDA, QuantizedCPU, QuantizedCUDA: fill_
+    CPU, CUDA: fill_
+    QuantizedCPU, QuantizedCUDA: fill_quantized_
     Meta: fill_meta_
 
 - func: fill_.Tensor(Tensor(a!) self, Tensor value) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   variants: function, method
   dispatch:
-    CPU, CUDA, QuantizedCPU, QuantizedCUDA: fill_
+    CPU, CUDA: fill_
+    QuantizedCPU, QuantizedCUDA: fill_quantized_
     Meta: fill_meta_
 
 - func: floor(Tensor self) -> Tensor
@@ -1987,6 +2111,8 @@
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: floor
+    SparseCPU, SparseCUDA: floor_sparse
+    SparseCsrCPU, SparseCsrCUDA: floor_sparse_csr
 
 - func: floor_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -1994,6 +2120,8 @@
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: floor_
+    SparseCPU, SparseCUDA: floor_sparse_
+    SparseCsrCPU, SparseCsrCUDA: floor_sparse_csr_
 
 - func: floor.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -2001,6 +2129,8 @@
   structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: floor_out
+    SparseCPU, SparseCUDA: floor_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: floor_sparse_csr_out
 
 - func: floor_divide(Tensor self, Tensor other) -> Tensor
   device_check: NoCheck   # TensorIterator
@@ -2108,10 +2238,13 @@
 
 - func: grid_sampler_2d(Tensor input, Tensor grid, int interpolation_mode, int padding_mode, bool align_corners) -> Tensor
   dispatch:
-    CPU: grid_sampler_2d_cpu
+    CPU, QuantizedCPU: grid_sampler_2d_cpu
     CUDA: grid_sampler_2d_cuda
 
-- func: grid_sampler_2d_backward(Tensor grad_output, Tensor input, Tensor grid, int interpolation_mode, int padding_mode, bool align_corners) -> (Tensor, Tensor)
+# `grid_sampler_2d_backward` takes in `output_mask` to optimize performance for
+# the case where `input` doesn't require gradient. Gradient for `grid` is always
+# computed (only `output_mask[0]` is checked by the implementations).
+- func: grid_sampler_2d_backward(Tensor grad_output, Tensor input, Tensor grid, int interpolation_mode, int padding_mode, bool align_corners, bool[2] output_mask) -> (Tensor, Tensor)
   dispatch:
     CPU: grid_sampler_2d_backward_cpu
     CUDA: grid_sampler_2d_backward_cuda
@@ -2229,6 +2362,8 @@
 
 - func: index_copy(Tensor self, int dim, Tensor index, Tensor source) -> Tensor
   variants: function, method
+  dispatch:
+    CompositeExplicitAutograd: index_copy
 
 - func: index_copy_.dimname(Tensor(a!) self, Dimname dim, Tensor index, Tensor source) -> Tensor(a!)
   variants: method
@@ -2250,6 +2385,8 @@
 - func: index_put(Tensor self, Tensor?[] indices, Tensor values, bool accumulate=False) -> Tensor
   device_check: NoCheck   # delegate to _index_put_impl_ after clone, which leverages TensorIterator
   variants: function, method
+  dispatch:
+    CompositeExplicitAutograd: index_put
 
 - func: _index_put_impl_(Tensor(a!) self, Tensor?[] indices, Tensor values, bool accumulate=False, bool unsafe=False) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -2269,12 +2406,6 @@
   dispatch:
     CompositeExplicitAutograd: inverse_out
 
-- func: _inverse_helper(Tensor self) -> Tensor
-  variants: function
-  dispatch:
-    CPU: _inverse_helper_cpu
-    CUDA: _inverse_helper_cuda
-
 - func: isclose(Tensor self, Tensor other, float rtol=1e-05, float atol=1e-08, bool equal_nan=False) -> Tensor
   variants: function, method
 
@@ -2315,6 +2446,7 @@
   dispatch:
     CPU, CUDA: isnan
     SparseCPU, SparseCUDA: isnan_sparse
+    SparseCsrCPU, SparseCsrCUDA: isnan_sparse_csr
 
 - func: is_distributed(Tensor self) -> bool
   variants: function, method
@@ -2338,6 +2470,11 @@
   device_guard: False
   manual_cpp_binding: True
 
+- func: _is_zerotensor(Tensor self) -> bool
+  variants: function, method
+  device_guard: False
+  manual_cpp_binding: True
+
 - func: is_neg(Tensor self) -> bool
   variants: function, method
   device_guard: False
@@ -2405,6 +2542,11 @@
     CUDA: layer_norm_cuda
     CompositeImplicitAutograd: math_native_layer_norm
 
+- func: _native_multi_head_self_attention(Tensor query, Tensor qkv_weight, Tensor qkv_bias, Tensor proj_weight, Tensor proj_bias, Tensor? mask=None) -> Tensor
+  dispatch:
+    CPU: multi_head_self_attention_cpu
+    CUDA: multi_head_self_attention_cuda
+
 - func: native_layer_norm_backward(Tensor grad_out, Tensor input, int[] normalized_shape, Tensor mean, Tensor rstd, Tensor? weight, Tensor? bias, bool[3] output_mask) -> (Tensor, Tensor, Tensor)
   dispatch:
     CPU: layer_norm_backward_cpu
@@ -2414,15 +2556,18 @@
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: nan_to_num
+    SparseCPU, SparseCUDA: nan_to_num_sparse
 
 - func: nan_to_num_(Tensor(a!) self, float? nan=None, float? posinf=None, float? neginf=None) -> Tensor(a!)
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: nan_to_num_
+    SparseCPU, SparseCUDA: nan_to_num_sparse_
 
 - func: nan_to_num.out(Tensor self, float? nan=None, float? posinf=None, float? neginf=None, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
     CPU, CUDA: nan_to_num_out
+    SparseCPU, SparseCUDA: nan_to_num_sparse_out
 
 - func: linear(Tensor input, Tensor weight, Tensor? bias=None) -> Tensor
   python_module: nn
@@ -2471,11 +2616,11 @@
 
 - func: ldexp.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
 
-- func: linspace(Scalar start, Scalar end, int? steps=None, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
+- func: linspace(Scalar start, Scalar end, int steps, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
 
-- func: linspace.out(Scalar start, Scalar end, int? steps=None, *, Tensor(a!) out) -> Tensor(a!)
+- func: linspace.out(Scalar start, Scalar end, int steps, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
-    CPU: linspace_cpu_out
+    CPU, Meta: linspace_out
     CUDA: linspace_cuda_out
 
 - func: log(Tensor self) -> Tensor
@@ -2499,6 +2644,8 @@
   device_check: NoCheck   # TensorIterator
   structured_delegate: log10.out
   variants: function, method
+  dispatch:
+    CompositeExplicitAutograd: log10
 
 - func: log10_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -2518,6 +2665,7 @@
   variants: function, method
   dispatch:
     SparseCPU, SparseCUDA: log1p_sparse
+    SparseCsrCPU, SparseCsrCUDA: log1p_sparse_csr
 
 - func: log1p_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -2525,6 +2673,7 @@
   variants: function, method
   dispatch:
     SparseCPU, SparseCUDA: log1p_sparse_
+    SparseCsrCPU, SparseCsrCUDA: log1p_sparse_csr_
 
 - func: log1p.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -2532,7 +2681,8 @@
   structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: log1p_out
-    SparseCPU, SparseCUDA: log1p_out_sparse
+    SparseCPU, SparseCUDA: log1p_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: log1p_sparse_csr_out
 
 - func: log2(Tensor self) -> Tensor
   device_check: NoCheck   # TensorIterator
@@ -2630,11 +2780,11 @@
   dispatch:
     CompositeExplicitAutograd: logdet
 
-- func: logspace(Scalar start, Scalar end, int? steps=None, float base=10.0, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
+- func: logspace(Scalar start, Scalar end, int steps, float base=10.0, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
 
-- func: logspace.out(Scalar start, Scalar end, int? steps=None, float base=10.0, *, Tensor(a!) out) -> Tensor(a!)
+- func: logspace.out(Scalar start, Scalar end, int steps, float base=10.0, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
-    CPU: logspace_cpu_out
+    CPU, Meta: logspace_out
     CUDA: logspace_cuda_out
 
 # log_softmax allows positional dtype, unlike most operators, because kwonly is BC-breaking when loading jit models.
@@ -2653,10 +2803,10 @@
     CPU: log_softmax_cpu_out
     CUDA: log_softmax_cuda_out
 
-- func: _log_softmax_backward_data(Tensor grad_output, Tensor output, int dim, Tensor self) -> Tensor
+- func: _log_softmax_backward_data(Tensor grad_output, Tensor output, int dim, ScalarType input_dtype) -> Tensor
   structured_delegate: _log_softmax_backward_data.out
 
-- func: _log_softmax_backward_data.out(Tensor grad_output, Tensor output, int dim, Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+- func: _log_softmax_backward_data.out(Tensor grad_output, Tensor output, int dim, ScalarType input_dtype, *, Tensor(a!) out) -> Tensor(a!)
   structured: True
   dispatch:
     CPU: log_softmax_backward_cpu_out
@@ -2722,11 +2872,11 @@
 # Alias to linalg.matrix_power
 - func: matrix_power.out(Tensor self, int n, *, Tensor(a!) out) -> Tensor(a!)
 
+# Alias to linalg.matrix_exp
 - func: matrix_exp(Tensor self) -> Tensor
   variants: function, method
-  dispatch:
-    CPU, CUDA: matrix_exp
 
+# This function should be deprecated in favor of differential_analytic_matrix_function in FunctionsManual.cpp
 - func: matrix_exp_backward(Tensor self, Tensor grad) -> Tensor
 
 # DEPRECATED: Use torch.aminmax instead
@@ -2760,12 +2910,16 @@
 
 - func: max.dim(Tensor self, int dim, bool keepdim=False) -> (Tensor values, Tensor indices)
   device_check: NoCheck   # TensorIterator
+  structured_delegate: max.dim_max
   variants: function, method
   dispatch:
-    CPU, CUDA, QuantizedCPU, QuantizedCUDA: max
+    QuantizedCPU, QuantizedCUDA: qmax
 
 - func: max.dim_max(Tensor self, int dim, bool keepdim=False, *, Tensor(a!) max, Tensor(b!) max_values) -> (Tensor(a!) values, Tensor(b!) indices)
   device_check: NoCheck   # TensorIterator
+  structured: True
+  precomputed:
+  - dim -> int dim
   dispatch:
     CPU, CUDA: max_out
 
@@ -2903,12 +3057,16 @@
 
 - func: min.dim(Tensor self, int dim, bool keepdim=False) -> (Tensor values, Tensor indices)
   device_check: NoCheck   # TensorIterator
+  structured_delegate: min.dim_min
   variants: function, method
   dispatch:
-    CPU, CUDA, QuantizedCPU, QuantizedCUDA: min
+    QuantizedCPU, QuantizedCUDA: qmin
 
 - func: min.dim_min(Tensor self, int dim, bool keepdim=False, *, Tensor(a!) min, Tensor(b!) min_indices) -> (Tensor(a!) values, Tensor(b!) indices)
   device_check: NoCheck   # TensorIterator
+  structured: True
+  precomputed:
+  - dim -> int dim
   dispatch:
     CPU, CUDA: min_out
 
@@ -2932,14 +3090,6 @@
   dispatch:
     CompositeExplicitAutograd: mkldnn_convolution
 
-- func: mkldnn_convolution_backward_input(int[] self_size, Tensor grad_output, Tensor weight, int[] padding, int[] stride, int[] dilation, int groups, bool bias_defined) -> Tensor
-
-- func: mkldnn_convolution_backward_weights(int[] weight_size, Tensor grad_output, Tensor self, int[] padding, int[] stride, int[] dilation, int groups, bool bias_defined) -> (Tensor, Tensor)
-
-- func: mkldnn_convolution_backward(Tensor self, Tensor grad_output, Tensor weight, int[] padding, int[] stride, int[] dilation, int groups, bool[3] output_mask) -> (Tensor, Tensor, Tensor)
-  dispatch:
-    CompositeExplicitAutograd: mkldnn_convolution_backward
-
 - func: miopen_batch_norm(Tensor input, Tensor weight, Tensor? bias, Tensor? running_mean, Tensor? running_var, bool training, float exponential_average_factor, float epsilon) -> (Tensor, Tensor, Tensor)
   dispatch:
     CUDA: miopen_batch_norm
@@ -2952,56 +3102,14 @@
   dispatch:
     CUDA: miopen_convolution
 
-- func: miopen_convolution_backward_input(int[] self_size, Tensor grad_output, Tensor weight, int[] padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic) -> Tensor
-  dispatch:
-    CUDA: miopen_convolution_backward_input
-
-- func: miopen_convolution_backward(Tensor self, Tensor grad_output, Tensor weight, int[] padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic, bool[3] output_mask) -> (Tensor, Tensor, Tensor)
-  dispatch:
-    CUDA: miopen_convolution_backward
-
-- func: miopen_convolution_backward_bias(Tensor grad_output) -> Tensor
-  dispatch:
-    CUDA: miopen_convolution_backward_bias
-
-- func: miopen_convolution_backward_weight(int[] weight_size, Tensor grad_output, Tensor self, int[] padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic) -> Tensor
-  dispatch:
-    CUDA: miopen_convolution_backward_weight
-
 - func: miopen_convolution_transpose(Tensor self, Tensor weight, Tensor? bias, int[] padding, int[] output_padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic) -> Tensor
   dispatch:
     CUDA: miopen_convolution_transpose
 
-# NB: output_padding not strictly needed here, but it's helpful for the float
-# backwards
-- func: miopen_convolution_transpose_backward(Tensor self, Tensor grad_output, Tensor weight, int[] padding, int[] output_padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic, bool[3] output_mask) -> (Tensor, Tensor, Tensor)
-  dispatch:
-    CUDA: miopen_convolution_transpose_backward
-
-- func: miopen_convolution_transpose_backward_input(Tensor grad_output, Tensor weight, int[] padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic) -> Tensor
-  dispatch:
-    CUDA: miopen_convolution_transpose_backward_input
-
-- func: miopen_convolution_transpose_backward_weight(int[] weight_size, Tensor grad_output, Tensor self, int[] padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic) -> Tensor
-  dispatch:
-    CUDA: miopen_convolution_transpose_backward_weight
-
 - func: miopen_depthwise_convolution(Tensor self, Tensor weight, Tensor? bias, int[] padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic) -> Tensor
   dispatch:
     CUDA: miopen_depthwise_convolution
 
-- func: miopen_depthwise_convolution_backward_input(int[] self_size, Tensor grad_output, Tensor weight, int[] padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic) -> Tensor
-  dispatch:
-    CUDA: miopen_depthwise_convolution_backward_input
-
-- func: miopen_depthwise_convolution_backward(Tensor self, Tensor grad_output, Tensor weight, int[] padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic, bool[3] output_mask) -> (Tensor, Tensor, Tensor)
-  dispatch:
-    CUDA: miopen_depthwise_convolution_backward
-
-- func: miopen_depthwise_convolution_backward_weight(int[] weight_size, Tensor grad_output, Tensor self, int[] padding, int[] stride, int[] dilation, int groups, bool benchmark, bool deterministic) -> Tensor
-  dispatch:
-    CUDA: miopen_depthwise_convolution_backward_weight
-
 - func: miopen_rnn(Tensor input, Tensor[] weight, int weight_stride0, Tensor hx, Tensor? cx, int mode, int hidden_size, int num_layers, bool batch_first, float dropout, bool train, bool bidirectional, int[] batch_sizes, Tensor? dropout_state) -> (Tensor, Tensor, Tensor, Tensor, Tensor)
   dispatch:
     CUDA: miopen_rnn
@@ -3014,7 +3122,8 @@
   structured_delegate: mm.out
   variants: function, method
   dispatch:
-    SparseCPU, SparseCUDA, SparseCsrCPU, SparseCsrCUDA: _sparse_mm
+    SparseCPU, SparseCUDA: _sparse_mm
+    SparseCsrCPU, SparseCsrCUDA: _sparse_csr_mm
 
 - func: mm.out(Tensor self, Tensor mat2, *, Tensor(a!) out) -> Tensor(a!)
   structured: True
@@ -3057,6 +3166,7 @@
   dispatch:
     SparseCPU, SparseCUDA: mul_sparse
     MkldnnCPU: mkldnn_mul
+    ZeroTensor: mul_zerotensor
 
 - func: mul_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -3107,8 +3217,8 @@
 - func: mv(Tensor self, Tensor vec) -> Tensor
   variants: function, method
   dispatch:
-    CPU, CUDA: mv
-    SparseCPU, SparseCUDA, SparseCsrCPU, SparseCsrCUDA: mv_sparse
+    CompositeExplicitAutograd: mv
+    SparseCPU, SparseCUDA: mv_sparse
 
 - func: mv.out(Tensor self, Tensor vec, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
@@ -3210,15 +3320,6 @@
   dispatch:
     CompositeExplicitAutograd: _nnpack_spatial_convolution
 
-- func: _nnpack_spatial_convolution_backward(Tensor input, Tensor grad_output, Tensor weight, int[2] padding, bool[3] output_mask) -> (Tensor, Tensor, Tensor)
-  variants: function
-
-- func: _nnpack_spatial_convolution_backward_input(Tensor input, Tensor grad_output, Tensor weight, int[2] padding) -> Tensor
-  variants: function
-
-- func: _nnpack_spatial_convolution_backward_weight(Tensor input, int[] weightsize, Tensor grad_output, int[2] padding) -> Tensor
-  variants: function
-
 - func: ones.names(int[] size, *, Dimname[]? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
   device_check: NoCheck
   device_guard: False
@@ -3286,6 +3387,21 @@
 - func: numpy_T(Tensor(a) self) -> Tensor(a)
   variants: method
 
+# Exposed on Python as an attribute 'H'
+- func: matrix_H(Tensor(a) self) -> Tensor(a)
+  variants: method
+
+# Exposed on Python as an attribute 'mT'
+- func: mT(Tensor(a) self) -> Tensor(a)
+  variants: method
+
+# Exposed on Python as an attribute 'mH'
+- func: mH(Tensor(a) self) -> Tensor(a)
+  variants: method
+
+- func: adjoint(Tensor(a) self) -> Tensor(a)
+  variants: function, method
+
 - func: pixel_shuffle(Tensor self, int upscale_factor) -> Tensor
 
 - func: pixel_unshuffle(Tensor self, int downscale_factor) -> Tensor
@@ -3295,6 +3411,11 @@
     CPU: channel_shuffle
     QuantizedCPU: channel_shuffle_quantized_cpu
 
+- func: native_channel_shuffle(Tensor self, int groups) -> Tensor
+  dispatch:
+    CPU: channel_shuffle_cpu
+    CompositeImplicitAutograd: math_channel_shuffle
+
 - func: is_pinned(Tensor self, Device? device=None) -> bool
   variants: method
   dispatch:
@@ -3321,15 +3442,18 @@
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: rad2deg
+    SparseCsrCPU, SparseCsrCUDA: rad2deg_sparse_csr
 
 - func: rad2deg_(Tensor(a!) self) -> Tensor(a!)
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: rad2deg_
+    SparseCsrCPU, SparseCsrCUDA: rad2deg_sparse_csr_
 
 - func: rad2deg.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
     CompositeExplicitAutograd: rad2deg_out
+    SparseCsrCPU, SparseCsrCUDA: rad2deg_sparse_csr_out
 
 - func: deg2rad(Tensor self) -> Tensor
   variants: function, method
@@ -3420,7 +3544,7 @@
 
 - func: range.out(Scalar start, Scalar end, Scalar step=1, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
-    CPU: range_cpu_out
+    CPU, Meta: range_out
     CUDA: range_cuda_out
 
 - func: ravel(Tensor(a) self) -> Tensor(a)
@@ -3449,6 +3573,7 @@
   variants: function, method
   dispatch:
     SparseCPU, SparseCUDA: neg_sparse
+    SparseCsrCPU, SparseCsrCUDA: neg_sparse_csr
 
 - func: neg_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -3456,6 +3581,7 @@
   variants: function, method
   dispatch:
     SparseCPU, SparseCUDA: neg_sparse_
+    SparseCsrCPU, SparseCsrCUDA: neg_sparse_csr_
 
 - func: neg.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -3464,6 +3590,7 @@
   dispatch:
     CPU, CUDA: neg_out
     SparseCPU, SparseCUDA: neg_out_sparse
+    SparseCsrCPU, SparseCsrCUDA: neg_sparse_csr_out
 
 # Alias for neg
 - func: negative(Tensor self) -> Tensor
@@ -3504,7 +3631,7 @@
   device_check: NoCheck
   device_guard: False
   dispatch:
-    CPU, CUDA, Meta, QuantizedCPU, QuantizedCUDA: _reshape_alias
+    CPU, CUDA, Meta, QuantizedCPU, QuantizedCUDA, ZeroTensor: _reshape_alias
     # We don't need to support mkldnn since this is handled explicitly by the reshape operator.
 
 - func: _mkldnn_reshape(Tensor self, int[] shape) -> Tensor
@@ -3522,11 +3649,17 @@
   device_check: NoCheck   # TensorIterator
   structured_delegate: round.out
   variants: function, method
+  dispatch:
+    SparseCPU, SparseCUDA: round_sparse
+    SparseCsrCPU, SparseCsrCUDA: round_sparse_csr
 
 - func: round_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   structured_delegate: round.out
   variants: function, method
+  dispatch:
+    SparseCPU, SparseCUDA: round_sparse_
+    SparseCsrCPU, SparseCsrCUDA: round_sparse_csr_
 
 - func: round.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -3535,6 +3668,26 @@
   dispatch:
     CPU: round_out
     CUDA: round_out
+    SparseCPU, SparseCUDA: round_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: round_sparse_csr_out
+
+- func: round.decimals(Tensor self, *, int decimals) -> Tensor
+  device_check: NoCheck   # TensorIterator
+  structured_delegate: round.decimals_out
+  variants: function, method
+
+- func: round_.decimals(Tensor(a!) self, *, int decimals) -> Tensor(a!)
+  device_check: NoCheck   # TensorIterator
+  structured_delegate: round.decimals_out
+  variants: function, method
+
+- func: round.decimals_out(Tensor self, *, int decimals, Tensor(a!) out) -> Tensor(a!)
+  device_check: NoCheck   # TensorIterator
+  structured: True
+  structured_inherits: TensorIteratorBase
+  dispatch:
+    CPU: round_decimals_out
+    CUDA: round_decimals_out
 
 - func: rrelu(Tensor self, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None) -> Tensor
   device_check: NoCheck   # TensorIterator
@@ -3591,6 +3744,7 @@
   python_module: nn
   dispatch:
     MkldnnCPU: mkldnn_gelu
+    QuantizedCPU: gelu_quantized_cpu
 
 - func: gelu_backward.grad_input(Tensor grad, Tensor self, *, Tensor(a!) grad_input) -> Tensor(a!)
   structured: True
@@ -3783,11 +3937,17 @@
   device_check: NoCheck   # TensorIterator
   structured_delegate: sin.out
   variants: function, method
+  dispatch:
+    SparseCsrCPU, SparseCsrCUDA: sin_sparse_csr
+    SparseCPU, SparseCUDA: sin_sparse
 
 - func: sin_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   structured_delegate: sin.out
   variants: function, method
+  dispatch:
+    SparseCsrCPU, SparseCsrCUDA: sin_sparse_csr_
+    SparseCPU, SparseCUDA: sin_sparse_
 
 - func: sin.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -3795,6 +3955,8 @@
   structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: sin_out
+    SparseCsrCPU, SparseCsrCUDA: sin_sparse_csr_out
+    SparseCPU, SparseCUDA: sin_sparse_out
 
 - func: sinc(Tensor self) -> Tensor
   structured_delegate: sinc.out
@@ -3814,11 +3976,17 @@
   device_check: NoCheck   # TensorIterator
   structured_delegate: sinh.out
   variants: function, method
+  dispatch:
+    SparseCPU, SparseCUDA: sinh_sparse
+    SparseCsrCPU, SparseCsrCUDA: sinh_sparse_csr
 
 - func: sinh_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   structured_delegate: sinh.out
   variants: function, method
+  dispatch:
+    SparseCPU, SparseCUDA: sinh_sparse_
+    SparseCsrCPU, SparseCsrCUDA: sinh_sparse_csr_
 
 - func: sinh.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -3826,6 +3994,8 @@
   structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: sinh_out
+    SparseCPU, SparseCUDA: sinh_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: sinh_sparse_csr_out
 
 # Returns a copy of this `Variable` that is detached from its autograd graph.
 # This method is OK to call if the `Variable` is a view.
@@ -3848,6 +4018,7 @@
 # this. If this `Variable` is a view, throws an `std::runtime_error()`.
 - func: detach_(Tensor(a!) self) -> Tensor(a!)
   variants: function, method
+  tags: inplace_view
   dispatch:
     CompositeExplicitAutograd: detach_
 
@@ -3876,6 +4047,27 @@
   dispatch:
     CompositeExplicitAutograd: slice_backward
 
+- func: slice_scatter(Tensor self, Tensor src, int dim=0, int? start=None, int? end=None, int step=1) -> Tensor
+  variants: function, method
+  device_check: NoCheck
+  device_guard: False
+  dispatch:
+    CompositeExplicitAutograd: slice_scatter
+
+- func: select_scatter(Tensor self, Tensor src, int dim, int index) -> Tensor
+  variants: function, method
+  device_check: NoCheck
+  device_guard: False
+  dispatch:
+    CompositeExplicitAutograd: select_scatter
+
+- func: diagonal_scatter(Tensor self, Tensor src, int offset=0, int dim1=0, int dim2=1) -> Tensor
+  variants: function, method
+  device_check: NoCheck
+  device_guard: False
+  dispatch:
+    CompositeExplicitAutograd: diagonal_scatter
+
 - func: slogdet(Tensor self) -> (Tensor sign, Tensor logabsdet)
   variants: function, method
   dispatch:
@@ -3902,10 +4094,10 @@
     CPU: softmax_cpu_out
     CUDA: softmax_cuda_out
 
-- func: _softmax_backward_data(Tensor grad_output, Tensor output, int dim, Tensor self) -> Tensor
+- func: _softmax_backward_data(Tensor grad_output, Tensor output, int dim, ScalarType input_dtype) -> Tensor
   structured_delegate: _softmax_backward_data.out
 
-- func: _softmax_backward_data.out(Tensor grad_output, Tensor output, int dim, Tensor self, *, Tensor(a!) grad_input) -> Tensor(a!)
+- func: _softmax_backward_data.out(Tensor grad_output, Tensor output, int dim, ScalarType input_dtype, *, Tensor(a!) grad_input) -> Tensor(a!)
   structured: True
   dispatch:
     CPU: softmax_backward_cpu_out
@@ -3918,7 +4110,7 @@
   dispatch:
     CompositeExplicitAutograd: unsafe_split
 
-- func: split.Tensor(Tensor(a) self, int split_size, int dim=0) -> Tensor(a)[]
+- func: split.Tensor(Tensor(a -> *) self, int split_size, int dim=0) -> Tensor(a)[]
   variants: function, method
   device_check: NoCheck
   device_guard: False
@@ -3932,29 +4124,29 @@
   dispatch:
     CompositeExplicitAutograd: unsafe_split_with_sizes
 
-- func: split_with_sizes(Tensor(a) self, int[] split_sizes, int dim=0) -> Tensor(a)[]
+- func: split_with_sizes(Tensor(a -> *) self, int[] split_sizes, int dim=0) -> Tensor(a)[]
   variants: function, method
   device_check: NoCheck
   device_guard: False
   dispatch:
     CompositeExplicitAutograd: split_with_sizes
 
-- func: hsplit.int(Tensor(a) self, int sections) -> Tensor(a)[]
+- func: hsplit.int(Tensor(a -> *) self, int sections) -> Tensor(a)[]
   variants: function, method
 
-- func: hsplit.array(Tensor(a) self, int[] indices) -> Tensor(a)[]
+- func: hsplit.array(Tensor(a -> *) self, int[] indices) -> Tensor(a)[]
   variants: function, method
 
-- func: vsplit.int(Tensor(a) self, int sections) -> Tensor(a)[]
+- func: vsplit.int(Tensor(a -> *) self, int sections) -> Tensor(a)[]
   variants: function, method
 
-- func: vsplit.array(Tensor(a) self, int[] indices) -> Tensor(a)[]
+- func: vsplit.array(Tensor(a -> *) self, int[] indices) -> Tensor(a)[]
   variants: function, method
 
-- func: dsplit.int(Tensor(a) self, int sections) -> Tensor(a)[]
+- func: dsplit.int(Tensor(a -> *) self, int sections) -> Tensor(a)[]
   variants: function, method
 
-- func: dsplit.array(Tensor(a) self, int[] indices) -> Tensor(a)[]
+- func: dsplit.array(Tensor(a -> *) self, int[] indices) -> Tensor(a)[]
   variants: function, method
 
 - func: squeeze(Tensor(a) self) -> Tensor(a)
@@ -3962,14 +4154,16 @@
   device_check: NoCheck
   device_guard: False
   dispatch:
-    CompositeExplicitAutograd: squeeze
+    CPU, CUDA: squeeze
+    QuantizedCPU, QuantizedCUDA: squeeze_quantized
 
 - func: squeeze.dim(Tensor(a) self, int dim) -> Tensor(a)
   variants: function, method
   device_check: NoCheck
   device_guard: False
   dispatch:
-    CompositeExplicitAutograd: squeeze
+    CPU, CUDA: squeeze
+    QuantizedCPU, QuantizedCUDA: squeeze_quantized
 
 - func: squeeze.dimname(Tensor(a) self, Dimname dim) -> Tensor(a)
   variants: function, method
@@ -3980,6 +4174,7 @@
   variants: method
   device_check: NoCheck
   device_guard: False
+  tags: inplace_view
   dispatch:
     CompositeExplicitAutograd: squeeze_
 
@@ -3987,6 +4182,7 @@
   variants: method
   device_check: NoCheck
   device_guard: False
+  tags: inplace_view
   dispatch:
     CompositeExplicitAutograd: squeeze_
 
@@ -3994,6 +4190,7 @@
   variants: method
   device_check: NoCheck
   device_guard: False
+  tags: inplace_view
 
 - func: sspaddmm(Tensor self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1) -> Tensor
   variants: function, method
@@ -4105,11 +4302,15 @@
   variants: function, method
   dispatch:
     SparseCPU, SparseCUDA: sqrt_sparse
+    SparseCsrCPU, SparseCsrCUDA: sqrt_sparse_csr
 
 - func: sqrt_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   structured_delegate: sqrt.out
   variants: function, method
+  dispatch:
+    SparseCPU, SparseCUDA: sqrt_sparse_
+    SparseCsrCPU, SparseCsrCUDA: sqrt_sparse_csr_
 
 - func: sqrt.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -4117,7 +4318,8 @@
   structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: sqrt_out
-    SparseCPU, SparseCUDA: sqrt_out_sparse
+    SparseCPU, SparseCUDA: sqrt_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: sqrt_sparse_csr_out
 
 - func: square(Tensor self) -> Tensor
   device_check: NoCheck   # TensorIterator
@@ -4225,6 +4427,7 @@
   device_check: NoCheck
   device_guard: False
   variants: method
+  tags: inplace_view
   dispatch:
     CompositeExplicitAutograd: t_
 
@@ -4232,11 +4435,17 @@
   device_check: NoCheck   # TensorIterator
   structured_delegate: tan.out
   variants: function, method
+  dispatch:
+    SparseCPU, SparseCUDA: tan_sparse
+    SparseCsrCPU, SparseCsrCUDA: tan_sparse_csr
 
 - func: tan_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   structured_delegate: tan.out
   variants: function, method
+  dispatch:
+    SparseCPU, SparseCUDA: tan_sparse_
+    SparseCsrCPU, SparseCsrCUDA: tan_sparse_csr_
 
 - func: tan.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -4244,6 +4453,8 @@
   structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: tan_out
+    SparseCPU, SparseCUDA: tan_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: tan_sparse_csr_out
 
 - func: tanh(Tensor self) -> Tensor
   device_check: NoCheck   # TensorIterator
@@ -4252,6 +4463,8 @@
   dispatch:
     QuantizedCPU: tanh_quantized_cpu
     MkldnnCPU: mkldnn_tanh
+    SparseCPU, SparseCUDA: tanh_sparse
+    SparseCsrCPU, SparseCsrCUDA: tanh_sparse_csr
 
 - func: tanh_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -4259,12 +4472,17 @@
   variants: function, method
   dispatch:
     MkldnnCPU: mkldnn_tanh_
+    SparseCPU, SparseCUDA: tanh_sparse_
+    SparseCsrCPU, SparseCsrCUDA: tanh_sparse_csr_
+
 - func: tanh.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: tanh_out
+    SparseCPU, SparseCUDA: tanh_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: tanh_sparse_csr_out
 
 - func: tensordot(Tensor self, Tensor other, int[] dims_self, int[] dims_other) -> Tensor
   variants: function
@@ -4331,6 +4549,7 @@
   variants: method
   device_check: NoCheck
   device_guard: False
+  tags: inplace_view
   dispatch:
     CompositeExplicitAutograd: transpose_
 
@@ -4388,6 +4607,8 @@
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: trunc
+    SparseCPU, SparseCUDA: trunc_sparse
+    SparseCsrCPU, SparseCsrCUDA: trunc_sparse_csr
 
 - func: trunc_(Tensor(a!) self) -> Tensor(a!)
   structured_delegate: trunc.out
@@ -4395,6 +4616,8 @@
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: trunc_
+    SparseCPU, SparseCUDA: trunc_sparse_
+    SparseCsrCPU, SparseCsrCUDA: trunc_sparse_csr_
 
 - func: trunc.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   structured: True
@@ -4402,6 +4625,8 @@
   device_check: NoCheck   # TensorIterator
   dispatch:
     CPU, CUDA: trunc_out
+    SparseCPU, SparseCUDA: trunc_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: trunc_sparse_csr_out
 
 # Alias for trunc
 - func: fix(Tensor self) -> Tensor
@@ -4461,12 +4686,15 @@
   device_check: NoCheck
   device_guard: False
   dispatch:
-    CompositeExplicitAutograd: unsqueeze
+    CPU, CUDA: unsqueeze
+    SparseCPU, SparseCUDA: unsqueeze_sparse
+    QuantizedCPU, QuantizedCUDA: unsqueeze_quantized
 
 - func: unsqueeze_(Tensor(a!) self, int dim) -> Tensor(a!)
   variants: method
   device_check: NoCheck
   device_guard: False
+  tags: inplace_view
   dispatch:
     CompositeExplicitAutograd: unsqueeze_
 
@@ -4586,6 +4814,11 @@
   device_check: NoCheck
   device_guard: False
 
+- func: _efficientzerotensor(int[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
+  dispatch:
+    CPU: _efficientzerotensor
+    CUDA: _efficientzerotensor_cuda
+
 - func: zeros(int[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
 
 - func: zeros.out(int[] size, *, Tensor(a!) out) -> Tensor(a!)
@@ -4655,12 +4888,15 @@
     SparseCUDA: _sparse_sum_backward_cuda
 
 - func: _sparse_softmax.int(Tensor self, int dim, ScalarType? dtype=None) -> Tensor
+  python_module: sparse
   variants: function
 
 - func: _sparse_softmax.Dimname(Tensor self, Dimname dim, *, ScalarType? dtype=None) -> Tensor
+  python_module: sparse
   variants: function
 
 - func: _sparse_softmax(Tensor self, int dim, bool half_to_float) -> Tensor
+  python_module: sparse
   dispatch:
     SparseCPU: softmax_sparse_cpu
     SparseCUDA: softmax_sparse_cuda
@@ -4671,12 +4907,15 @@
     SparseCUDA: softmax_backward_sparse_cuda
 
 - func: _sparse_log_softmax.int(Tensor self, int dim, ScalarType? dtype=None) -> Tensor
+  python_module: sparse
   variants: function
 
 - func: _sparse_log_softmax.Dimname(Tensor self, Dimname dim, *, ScalarType? dtype=None) -> Tensor
+  python_module: sparse
   variants: function
 
 - func: _sparse_log_softmax(Tensor self, int dim, bool half_to_float) -> Tensor
+  python_module: sparse
   dispatch:
     SparseCPU: log_softmax_sparse_cpu
     SparseCUDA: log_softmax_sparse_cuda
@@ -4774,6 +5013,7 @@
   dispatch:
     CompositeExplicitAutograd: clone
     SparseCPU, SparseCUDA: clone_sparse
+    SparseCsrCPU, SparseCsrCUDA: clone_sparse_csr
     MkldnnCPU: mkldnn_clone
     QuantizedCPU, QuantizedCUDA: quantized_clone
 
@@ -4886,9 +5126,20 @@
 # Functionally the same as addmm, but we give it a different derivative formula
 # that doesn't propagate gradients to non-present entries on sparse.
 - func: _sparse_addmm(Tensor self, Tensor sparse, Tensor dense, *, Scalar beta=1, Scalar alpha=1) -> Tensor
+  python_module: sparse
   dispatch:
     CompositeExplicitAutograd: _sparse_addmm
 
+- func: sparse_sampled_addmm.out(Tensor self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1, Tensor(a!) out) -> Tensor(a!)
+  python_module: sparse
+  dispatch:
+    SparseCsrCUDA: sparse_sampled_addmm_out_sparse_csr_cuda
+
+- func: sparse_sampled_addmm(Tensor self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1) -> Tensor
+  python_module: sparse
+  dispatch:
+    SparseCsrCUDA: sparse_sampled_addmm_sparse_csr_cuda
+
 - func: addmm.out(Tensor self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1, Tensor(a!) out) -> Tensor(a!)
   structured: True
   dispatch:
@@ -4896,8 +5147,8 @@
     CUDA: addmm_out_cuda
     SparseCPU: addmm_out_sparse_dense_cpu
     SparseCUDA: addmm_out_sparse_dense_cuda
-    SparseCsrCPU: addmm_out_sparse_csr_dense_cpu
-    SparseCsrCUDA: addmm_out_sparse_csr_dense_cuda
+    SparseCsrCPU: addmm_out_sparse_csr_cpu
+    SparseCsrCUDA: addmm_out_sparse_csr_cuda
 
 - func: addmm(Tensor self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1) -> Tensor
   structured_delegate: addmm.out
@@ -5209,12 +5460,12 @@
   dispatch:
     SparseCPU, SparseCUDA: copy_sparse_
 
-- func: unbind.int(Tensor(a) self, int dim=0) -> Tensor(a)[]
+- func: unbind.int(Tensor(a -> *) self, int dim=0) -> Tensor(a)[]
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: unbind
 
-- func: unbind.Dimname(Tensor(a) self, Dimname dim) -> Tensor(a)[]
+- func: unbind.Dimname(Tensor(a -> *) self, Dimname dim) -> Tensor(a)[]
   variants: function, method
 
 - func: to_sparse.sparse_dim(Tensor self, int sparse_dim) -> Tensor
@@ -5246,6 +5497,11 @@
 
 - func: to_mkldnn_backward(Tensor grad, Tensor input) -> Tensor
 
+- func: quantize_per_tensor_dynamic(Tensor self, ScalarType dtype, bool reduce_range) -> Tensor
+  variants: function
+  dispatch:
+    CPU, CUDA: quantize_per_tensor_dynamic
+
 - func: quantize_per_tensor(Tensor self, float scale, int zero_point, ScalarType dtype) -> Tensor
   variants: function
   dispatch:
@@ -5269,7 +5525,7 @@
 - func: dequantize.self(Tensor self) -> Tensor
   variants: function, method
   dispatch:
-    CPU: dequantize_cpu
+    CPU, CUDA: dequantize_cpu_or_cuda
     QuantizedCPU, QuantizedCUDA: dequantize_quantized
 
 - func: dequantize.tensors(Tensor[] tensors) -> Tensor[]
@@ -5391,6 +5647,14 @@
 - func: choose_qparams_optimized(Tensor input, int numel, int n_bins, float ratio, int bit_width) -> (Tensor, Tensor)
   variants: function
 
+- func: _autocast_to_reduced_precision(Tensor(a) self, bool cuda_enabled, bool cpu_enabled, ScalarType cuda_dtype, ScalarType cpu_dtype) -> Tensor(a)
+  variants: method
+  device_guard: False
+
+- func: _autocast_to_full_precision(Tensor(a) self, bool cuda_enabled, bool cpu_enabled) -> Tensor(a)
+  variants: method
+  device_guard: False
+
 - func: _to_copy(Tensor self, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None, bool non_blocking=False, MemoryFormat? memory_format=None) -> Tensor
   device_check: NoCheck
   device_guard: False
@@ -5589,6 +5853,8 @@
 - func: masked_fill.Scalar(Tensor self, Tensor mask, Scalar value) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
+  dispatch:
+    CompositeExplicitAutograd: masked_fill
 
 - func: masked_fill_.Tensor(Tensor(a!) self, Tensor mask, Tensor value) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -5600,6 +5866,8 @@
 - func: masked_fill.Tensor(Tensor self, Tensor mask, Tensor value) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
+  dispatch:
+    CompositeExplicitAutograd: masked_fill
 
 - func: masked_scatter_(Tensor(a!) self, Tensor mask, Tensor source) -> Tensor(a!)
   variants: method
@@ -5609,13 +5877,20 @@
 
 - func: masked_scatter(Tensor self, Tensor mask, Tensor source) -> Tensor
   variants: function, method
+  dispatch:
+    CompositeExplicitAutograd: masked_scatter
+
+- func: _masked_softmax(Tensor self, Tensor mask) -> Tensor
+  dispatch:
+    CUDA: masked_softmax_cuda
+    CPU: masked_softmax_cpu
 
 - func: view(Tensor(a) self, int[] size) -> Tensor(a)
   variants: method
   device_check: NoCheck
   device_guard: False
   dispatch:
-    CPU, CUDA, Meta, QuantizedCPU, QuantizedCUDA: view
+    ZeroTensor, CPU, CUDA, Meta, QuantizedCPU, QuantizedCUDA: view
     MkldnnCPU: mkldnn_view
 
 # Warning: If you want to change the name or overload name of this
@@ -5639,19 +5914,21 @@
 - func: put(Tensor self, Tensor index, Tensor source, bool accumulate=False) -> Tensor
   variants: function, method
 
-- func: index_add_(Tensor(a!) self, int dim, Tensor index, Tensor source) -> Tensor(a!)
-  variants: method
-
-- func: index_add_.alpha(Tensor(a!) self, int dim, Tensor index, Tensor source, *, Scalar alpha) -> Tensor(a!)
-  variants: method
+- func: index_add.out(Tensor self, int dim, Tensor index, Tensor source, *, Scalar alpha=1, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  variants: function
+  precomputed:
+  - dim -> int dim
   dispatch:
-    CPU: index_add_cpu_
-    CUDA: index_add_cuda_
+    CPU: index_add_cpu_out
+    CUDA: index_add_cuda_out
 
-- func: index_add(Tensor self, int dim, Tensor index, Tensor source) -> Tensor
-  variants: function, method
+- func: index_add_(Tensor(a!) self, int dim, Tensor index, Tensor source, *, Scalar alpha=1) -> Tensor(a!)
+  structured_delegate: index_add.out
+  variants: method
 
-- func: index_add.alpha(Tensor self, int dim, Tensor index, Tensor source, *, Scalar alpha) -> Tensor
+- func: index_add(Tensor self, int dim, Tensor index, Tensor source, *, Scalar alpha=1) -> Tensor
+  structured_delegate: index_add.out
   variants: function, method
 
 - func: index_add.dimname(Tensor self, Dimname dim, Tensor index, Tensor source, *, Scalar alpha=1) -> Tensor
@@ -5667,6 +5944,8 @@
 - func: index_fill.int_Scalar(Tensor self, int dim, Tensor index, Scalar value) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
+  dispatch:
+    CompositeExplicitAutograd: index_fill
 
 - func: index_fill_.int_Tensor(Tensor(a!) self, int dim, Tensor index, Tensor value) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -5677,6 +5956,8 @@
 - func: index_fill.int_Tensor(Tensor self, int dim, Tensor index, Tensor value) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
+  dispatch:
+    CompositeExplicitAutograd: index_fill
 
 - func: index_fill_.Dimname_Scalar(Tensor(a!) self, Dimname dim, Tensor index, Scalar value) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -5773,6 +6054,11 @@
 - func: scatter_add.dimname(Tensor self, Dimname dim, Tensor index, Tensor src) -> Tensor
   variants: function, method
 
+- func: scatter_reduce.two(Tensor self, int dim, Tensor index, str reduce, *, int? output_size=None) -> Tensor
+  variants: function, method
+  dispatch:
+    CPU: scatter_reduce_two_cpu
+
 - func: eq_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)
   structured_delegate: eq.Scalar_out
   device_check: NoCheck   # TensorIterator
@@ -6064,16 +6350,12 @@
     CPU, CUDA: bitwise_right_shift
 
 - func: tril_(Tensor(a!) self, int diagonal=0) -> Tensor(a!)
+  structured_delegate: tril.out
   variants: method
-  dispatch:
-    CPU: tril_cpu_
-    CUDA: tril_cuda_
 
 - func: triu_(Tensor(a!) self, int diagonal=0) -> Tensor(a!)
+  structured_delegate: triu.out
   variants: method
-  dispatch:
-    CPU: triu_cpu_
-    CUDA: triu_cuda_
 
 - func: digamma_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -6083,16 +6365,12 @@
 - func: lerp_.Scalar(Tensor(a!) self, Tensor end, Scalar weight) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   variants: method
-  dispatch:
-    CPU: lerp_cpu_scalar_
-    CUDA: lerp_cuda_scalar_
+  structured_delegate: lerp.Scalar_out
 
 - func: lerp_.Tensor(Tensor(a!) self, Tensor end, Tensor weight) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   variants: method
-  dispatch:
-    CPU: lerp_cpu_tensor_
-    CUDA: lerp_cuda_tensor_
+  structured_delegate: lerp.Tensor_out
 
 - func: addbmm_(Tensor(a!) self, Tensor batch1, Tensor batch2, *, Scalar beta=1, Scalar alpha=1) -> Tensor(a!)
   variants: method
@@ -6178,33 +6456,29 @@
   device_guard: False
 
 - func: cross.out(Tensor self, Tensor other, int? dim=None, *, Tensor(a!) out) -> Tensor(a!)
-  dispatch:
-    CPU, CUDA: cross_out
 
 - func: cross(Tensor self, Tensor other, int? dim=None) -> Tensor
   variants: method, function
-  dispatch:
-    CPU, CUDA: cross
 
 - func: triu.out(Tensor self, int diagonal=0, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
   dispatch:
-    CPU: triu_cpu_out
-    CUDA: triu_cuda_out
+    CPU: triu_cpu
+    CUDA: triu_cuda
 
 - func: triu(Tensor self, int diagonal=0) -> Tensor
+  structured_delegate: triu.out
   variants: method, function
-  dispatch:
-    CompositeExplicitAutograd: triu
 
 - func: tril.out(Tensor self, int diagonal=0, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
   dispatch:
-    CPU: tril_cpu_out
-    CUDA: tril_cuda_out
+    CPU: tril_cpu
+    CUDA: tril_cuda
 
 - func: tril(Tensor self, int diagonal=0) -> Tensor
+  structured_delegate: tril.out
   variants: method, function
-  dispatch:
-    CompositeExplicitAutograd: tril
 
 - func: tril_indices(int row, int col, int offset=0, *, ScalarType? dtype=long, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
   dispatch:
@@ -6584,7 +6858,8 @@
 - func: index_select(Tensor self, int dim, Tensor index) -> Tensor
   variants: method, function
   dispatch:
-    CPU, QuantizedCPU: index_select_cpu_
+    CPU: index_select_cpu_
+    QuantizedCPU: index_select_quantized_cpu_
     CUDA, QuantizedCUDA: index_select_cuda
     SparseCPU: index_select_sparse
     SparseCUDA: index_select_sparse
@@ -6629,6 +6904,9 @@
 - func: nonzero_numpy(Tensor self) -> Tensor[]
   variants: method, function
 
+- func: argwhere(Tensor self) -> Tensor
+  variants: method, function
+
 - func: gather.out(Tensor self, int dim, Tensor index, *, bool sparse_grad=False, Tensor(a!) out) -> Tensor(a!)
   structured: True
   dispatch:
@@ -6699,13 +6977,30 @@
     CUDA: legacy_lstsq_cuda
 
 - func: triangular_solve.X(Tensor self, Tensor A, bool upper=True, bool transpose=False, bool unitriangular=False, *, Tensor(a!) X, Tensor(b!) M) -> (Tensor(a!) solution, Tensor(b!) cloned_coefficient)
+  structured: True
   dispatch:
     CPU, CUDA: triangular_solve_out
+    SparseCsrCPU: triangular_solve_out_sparse_csr_cpu
+    SparseCsrCUDA: triangular_solve_out_sparse_csr_cuda
 
 - func: triangular_solve(Tensor self, Tensor A, bool upper=True, bool transpose=False, bool unitriangular=False) -> (Tensor solution, Tensor cloned_coefficient)
+  structured_delegate: triangular_solve.X
+  variants: method, function
+
+- func: _linalg_check_errors(Tensor info, str api_name, *, bool is_matrix) -> ()
+  dispatch:
+    CompositeExplicitAutograd: _linalg_check_errors
+
+- func: linalg_solve_triangular.out(Tensor self, Tensor B, *, bool upper, bool left=True, bool unitriangular=False, Tensor(a!) out) -> Tensor(a!)
+  python_module: linalg
+  dispatch:
+    CPU, CUDA: linalg_solve_triangular_out
+
+- func: linalg_solve_triangular(Tensor self, Tensor B, *, bool upper, bool left=True, bool unitriangular=False) -> Tensor
+  python_module: linalg
   variants: method, function
   dispatch:
-    CPU, CUDA: triangular_solve
+    CPU, CUDA: linalg_solve_triangular
 
 - func: symeig.e(Tensor self, bool eigenvectors=False, bool upper=True, *, Tensor(a!) e, Tensor(b!) V) -> (Tensor(a!) eigenvalues, Tensor(b!) eigenvectors)
   dispatch:
@@ -6736,12 +7031,6 @@
 - func: svd(Tensor self, bool some=True, bool compute_uv=True) -> (Tensor U, Tensor S, Tensor V)
   variants: method, function
 
-- func: _svd_helper(Tensor self, bool some, bool compute_uv) -> (Tensor U, Tensor S, Tensor V)
-  variants: function
-  dispatch:
-    CPU: _svd_helper_cpu
-    CUDA: _svd_helper_cuda
-
 # swapaxes, alias for transpose
 - func: swapaxes(Tensor(a) self, int axis0, int axis1) -> Tensor(a)
   variants: function, method
@@ -6752,6 +7041,7 @@
   variants: method
   device_check: NoCheck
   device_guard: False
+  tags: inplace_view
 
 # swapdims, alias for transpose
 - func: swapdims(Tensor(a) self, int dim0, int dim1) -> Tensor(a)
@@ -6763,6 +7053,7 @@
   variants: method
   device_check: NoCheck
   device_guard: False
+  tags: inplace_view
 
 - func: cholesky.out(Tensor self, bool upper=False, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
@@ -6843,8 +7134,6 @@
 
 - func: _lu_with_info(Tensor self, bool pivot=True, bool check_errors=True) -> (Tensor LU, Tensor pivots, Tensor info)
   variants: function
-  dispatch:
-    CPU, CUDA: _lu_with_info
 
 - func: lu_solve.out(Tensor self, Tensor LU_data, Tensor LU_pivots, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
@@ -6926,11 +7215,17 @@
   device_check: NoCheck   # TensorIterator
   structured_delegate: erfinv.out
   variants: method, function
+  dispatch:
+    SparseCPU, SparseCUDA: erfinv_sparse
+    SparseCsrCPU, SparseCsrCUDA: erfinv_sparse_csr
 
 - func: erfinv_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   structured_delegate: erfinv.out
   variants: method
+  dispatch:
+    SparseCPU, SparseCUDA: erfinv_sparse_
+    SparseCsrCPU, SparseCsrCUDA: erfinv_sparse_csr_
 
 - func: erfinv.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -6938,6 +7233,8 @@
   structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: erfinv_out
+    SparseCPU, SparseCUDA: erfinv_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: erfinv_sparse_csr_out
 
 - func: i0(Tensor self) -> Tensor
   structured_delegate: i0.out
@@ -6959,6 +7256,8 @@
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: sign
+    SparseCPU, SparseCUDA: sign_sparse
+    SparseCsrCPU, SparseCsrCUDA: sign_sparse_csr
 
 - func: sign_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -6966,6 +7265,8 @@
   variants: method
   dispatch:
     CompositeExplicitAutograd: sign_
+    SparseCPU, SparseCUDA: sign_sparse_
+    SparseCsrCPU, SparseCsrCUDA: sign_sparse_csr_
 
 - func: sign.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -6973,10 +7274,15 @@
   structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: sign_out
+    SparseCPU, SparseCUDA: sign_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: sign_sparse_csr_out
 
 - func: signbit(Tensor self) -> Tensor
   variants: function, method
   structured_delegate: signbit.out
+  dispatch:
+    SparseCPU, SparseCUDA: signbit_sparse
+    SparseCsrCPU, SparseCsrCUDA: signbit_sparse_csr
 
 - func: signbit.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   structured: True
@@ -6984,6 +7290,8 @@
   dispatch:
     CPU: signbit_out
     CUDA: signbit_out
+    SparseCPU, SparseCUDA: signbit_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: signbit_sparse_csr_out
 
 - func: dist(Tensor self, Tensor other, Scalar p=2) -> Tensor
   device_check: NoCheck   # TensorIterator
@@ -7008,31 +7316,39 @@
   structured_delegate: atan2.out
   variants: method, function
 
+# arctan2, alias of atan2
+- func: arctan2(Tensor self, Tensor other) -> Tensor
+  variants: method, function
+
+- func: arctan2.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
+  device_check: NoCheck   # TensorIterator
+
+- func: arctan2_(Tensor(a!) self, Tensor other) -> Tensor(a!)
+  variants: method
+
 - func: lerp.Scalar_out(Tensor self, Tensor end, Scalar weight, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
+  structured: True
+  structured_inherits: TensorIteratorBase
   dispatch:
-    CPU: lerp_cpu_scalar_out
-    CUDA: lerp_cuda_scalar_out
+    CPU, CUDA: lerp_Scalar
 
 - func: lerp.Tensor_out(Tensor self, Tensor end, Tensor weight, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
+  structured: True
+  structured_inherits: TensorIteratorBase
   dispatch:
-    CPU: lerp_cpu_tensor_out
-    CUDA: lerp_cuda_tensor_out
+    CPU, CUDA: lerp_Tensor
 
 - func: lerp.Scalar(Tensor self, Tensor end, Scalar weight) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: method, function
-  dispatch:
-    CPU: lerp_cpu_scalar
-    CUDA: lerp_cuda_scalar
+  structured_delegate: lerp.Scalar_out
 
 - func: lerp.Tensor(Tensor self, Tensor end, Tensor weight) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: method, function
-  dispatch:
-    CPU: lerp_cpu_tensor
-    CUDA: lerp_cuda_tensor
+  structured_delegate: lerp.Tensor_out
 
 - func: histc.out(Tensor self, int bins=100, Scalar min=0, Scalar max=0, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
@@ -7063,6 +7379,18 @@
   dispatch:
     CPU: histogram_cpu
 
+- func: _histogramdd_bin_edges(Tensor self, int[] bins, *, float[]? range=None, Tensor? weight=None, bool density=False) -> Tensor[]
+  dispatch:
+    CPU: histogramdd_bin_edges_cpu
+
+- func: _histogramdd_from_bin_cts(Tensor self, int[] bins, *, float[]? range=None, Tensor? weight=None, bool density=False) -> Tensor
+  dispatch:
+    CPU: histogramdd_cpu
+
+- func: _histogramdd_from_bin_tensors(Tensor self, Tensor[] bins, *, Tensor? weight=None, bool density=False) -> Tensor
+  dispatch:
+    CPU: histogramdd_cpu
+
 - func: fmod.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   dispatch:
@@ -7275,48 +7603,25 @@
   device_check: NoCheck   # TensorIterator
   variants: method, function
 
-# The following quantile signatures are DEPRECATED in favor of the new ones with the interpolation kwarg.
-- func: quantile.scalar_out(Tensor self, float q, int? dim=None, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
-
-- func: quantile.scalar(Tensor self, float q, int? dim=None, bool keepdim=False) -> Tensor
+- func: quantile(Tensor self, Tensor q, int? dim=None, bool keepdim=False, *, str interpolation='linear') -> Tensor
   variants: method, function
 
-- func: quantile.out(Tensor self, Tensor q, int? dim=None, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
+- func: quantile.out(Tensor self, Tensor q, int? dim=None, bool keepdim=False, *, str interpolation='linear', Tensor(a!) out) -> Tensor(a!)
 
-- func: quantile(Tensor self, Tensor q, int? dim=None, bool keepdim=False) -> Tensor
+- func: quantile.scalar(Tensor self, float q, int? dim=None, bool keepdim=False, *, str interpolation='linear') -> Tensor
   variants: method, function
 
-- func: nanquantile.scalar_out(Tensor self, float q, int? dim=None, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
+- func: quantile.scalar_out(Tensor self, float q, int? dim=None, bool keepdim=False, *, str interpolation='linear', Tensor(a!) out) -> Tensor(a!)
 
-- func: nanquantile.scalar(Tensor self, float q, int? dim=None, bool keepdim=False) -> Tensor
+- func: nanquantile(Tensor self, Tensor q, int? dim=None, bool keepdim=False, *, str interpolation='linear') -> Tensor
   variants: method, function
 
-- func: nanquantile.out(Tensor self, Tensor q, int? dim=None, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
+- func: nanquantile.out(Tensor self, Tensor q, int? dim=None, bool keepdim=False, *, str interpolation='linear', Tensor(a!) out) -> Tensor(a!)
 
-- func: nanquantile(Tensor self, Tensor q, int? dim=None, bool keepdim=False) -> Tensor
+- func: nanquantile.scalar(Tensor self, float q, int? dim=None, bool keepdim=False, *, str interpolation='linear') -> Tensor
   variants: method, function
 
-# To keep backward and forward compatibility, and to avoid ambiguity with the original signatures, dim, keepdim and interpolation
-# parameters are required for now. Once the deprecated signatures are removed they will be made optional.
-- func: quantile.new_scalar_out(Tensor self, float q, int? dim, bool keepdim, *, str interpolation, Tensor(a!) out) -> Tensor(a!)
-
-- func: quantile.new_scalar(Tensor self, float q, int? dim, bool keepdim, *, str interpolation) -> Tensor
-  variants: method, function
-
-- func: quantile.new_out(Tensor self, Tensor q, int? dim, bool keepdim, *, str interpolation, Tensor(a!) out) -> Tensor(a!)
-
-- func: quantile.new(Tensor self, Tensor q, int? dim, bool keepdim, *, str interpolation) -> Tensor
-  variants: method, function
-
-- func: nanquantile.new_scalar_out(Tensor self, float q, int? dim, bool keepdim, *, str interpolation, Tensor(a!) out) -> Tensor(a!)
-
-- func: nanquantile.new_scalar(Tensor self, float q, int? dim, bool keepdim, *, str interpolation) -> Tensor
-  variants: method, function
-
-- func: nanquantile.new_out(Tensor self, Tensor q, int? dim, bool keepdim, *, str interpolation, Tensor(a!) out) -> Tensor(a!)
-
-- func: nanquantile.new(Tensor self, Tensor q, int? dim, bool keepdim, *, str interpolation) -> Tensor
-  variants: method, function
+- func: nanquantile.scalar_out(Tensor self, float q, int? dim=None, bool keepdim=False, *, str interpolation='linear', Tensor(a!) out) -> Tensor(a!)
 
 - func: sort.values(Tensor self, int dim=-1, bool descending=False, *, Tensor(a!) values, Tensor(b!) indices) -> (Tensor(a!) values, Tensor(b!) indices)
   device_check: NoCheck   # TensorIterator
@@ -7511,6 +7816,7 @@
   dispatch:
     CPU, CUDA: normal_
     Meta: normal_meta_
+    SparseCsrCPU, SparseCsrCUDA: normal_sparse_csr_
 
 - func: normal.Tensor_float_out(Tensor mean, float std=1, *, Generator? generator=None, Tensor(a!) out) -> Tensor(a!)
   dispatch:
@@ -8209,6 +8515,13 @@
     CPU: foreach_tensor_minimum_slow
     CUDA: foreach_tensor_minimum_cuda
 
+- func: _foreach_norm.Scalar(Tensor[] tensors, Scalar ord=2) -> Tensor[]
+  device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
+  variants: function
+  dispatch:
+    CPU: foreach_tensor_norm_slow
+    CUDA: foreach_tensor_norm_cuda
+
 - func: bucketize.Tensor(Tensor self, Tensor boundaries, *, bool out_int32=False, bool right=False) -> Tensor
   dispatch:
     CPU: bucketize_cpu
@@ -8224,17 +8537,27 @@
     CPU: bucketize_cpu
     CUDA: bucketize_cuda
 
-- func: searchsorted.Tensor(Tensor sorted_sequence, Tensor self, *, bool out_int32=False, bool right=False) -> Tensor
+- func: searchsorted.Tensor(Tensor sorted_sequence, Tensor self, *, bool out_int32=False, bool right=False, str? side=None, Tensor? sorter=None) -> Tensor
   dispatch:
     CPU: searchsorted_cpu
     CUDA: searchsorted_cuda
 
-- func: searchsorted.Tensor_out(Tensor sorted_sequence, Tensor self, *, bool out_int32=False, bool right=False, Tensor(a!) out) -> Tensor(a!)
+# [Note about _torch_cuda_cu_linker_symbol_op and torch_cuda_cu]
+# This is a DUMMY function to force the linking against torch_cuda_cu on Windows.
+# Otherwise, the Windows linker will optimize and not include torch_cuda_cu even when we
+# want it to be included. This is similar to what we do with warp_size for torch_cuda_cpp,
+# described as the solution to this issue: https://github.com/pytorch/pytorch/issues/31611
+# This op should NOT be used or exposed or edited or else Windows builds (with BUILD_SPLIT_CUDA) will break.
+- func: _torch_cuda_cu_linker_symbol_op(Tensor self) -> Tensor
+  dispatch:
+    CUDA: _torch_cuda_cu_linker_symbol_op_cuda
+
+- func: searchsorted.Tensor_out(Tensor sorted_sequence, Tensor self, *, bool out_int32=False, bool right=False, str? side=None, Tensor? sorter=None, Tensor(a!) out) -> Tensor(a!)
   dispatch:
     CPU: searchsorted_out_cpu
     CUDA: searchsorted_out_cuda
 
-- func: searchsorted.Scalar(Tensor sorted_sequence, Scalar self, *, bool out_int32=False, bool right=False) -> Tensor
+- func: searchsorted.Scalar(Tensor sorted_sequence, Scalar self, *, bool out_int32=False, bool right=False, str? side=None, Tensor? sorter=None) -> Tensor
   dispatch:
     CPU: searchsorted_cpu
     CUDA: searchsorted_cuda
@@ -8248,6 +8571,15 @@
     CPU: _convert_indices_from_coo_to_csr_structured_cpu
     CUDA: _convert_indices_from_coo_to_csr_structured_cuda
 
+- func: _convert_indices_from_csr_to_coo(Tensor crow_indices, Tensor col_indices, *, bool out_int32=False, bool transpose=False) -> Tensor
+  structured_delegate: _convert_indices_from_csr_to_coo.out
+
+- func: _convert_indices_from_csr_to_coo.out(Tensor crow_indices, Tensor col_indices, *, bool out_int32=False, bool transpose=False, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  dispatch:
+    CPU: _convert_indices_from_csr_to_coo_structured_cpu
+    CUDA: _convert_indices_from_csr_to_coo_structured_cuda
+
 ## NN wrappers
 
 - func: mse_loss.out(Tensor self, Tensor target, int reduction=Mean, *, Tensor(a!) out) -> Tensor(a!)
@@ -8409,16 +8741,16 @@
 
 - func: smooth_l1_loss.out(Tensor self, Tensor target, int reduction=Mean, float beta=1.0, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
+  structured: True
+  structured_inherits: TensorIteratorBase
   python_module: nn
   dispatch:
-    CPU: smooth_l1_loss_out
-    CUDA: smooth_l1_loss_out
+    CPU, CUDA: smooth_l1_loss_out
 
 - func: smooth_l1_loss(Tensor self, Tensor target, int reduction=Mean, float beta=1.0) -> Tensor
   device_check: NoCheck   # TensorIterator
+  structured_delegate: smooth_l1_loss.out
   python_module: nn
-  dispatch:
-    CPU, CUDA: smooth_l1_loss
 
 - func: smooth_l1_loss_backward.grad_input(Tensor grad_output, Tensor self, Tensor target, int reduction, float beta, *, Tensor(a!) grad_input) -> Tensor(a!)
   python_module: nn
@@ -8533,6 +8865,7 @@
   python_module: nn
   dispatch:
     CPU, CUDA: hardsigmoid_out
+    QuantizedCPU: hardsigmoid_out_quantized_cpu
 
 - func: hardsigmoid(Tensor self) -> Tensor
   structured_delegate: hardsigmoid.out
@@ -8715,14 +9048,14 @@
   device_check: NoCheck   # TensorIterator
   python_module: nn
 
-- func: softplus_backward.grad_input(Tensor grad_output, Tensor self, Scalar beta, Scalar threshold, Tensor output, *, Tensor(a!) grad_input) -> Tensor(a!)
+- func: softplus_backward.grad_input(Tensor grad_output, Tensor self, Scalar beta, Scalar threshold, *, Tensor(a!) grad_input) -> Tensor(a!)
   structured: True
   structured_inherits: TensorIteratorBase
   python_module: nn
   dispatch:
     CPU, CUDA: softplus_backward_out
 
-- func: softplus_backward(Tensor grad_output, Tensor self, Scalar beta, Scalar threshold, Tensor output) -> Tensor
+- func: softplus_backward(Tensor grad_output, Tensor self, Scalar beta, Scalar threshold) -> Tensor
   structured_delegate: softplus_backward.grad_input
   python_module: nn
 
@@ -8933,19 +9266,22 @@
 
 - func: fractional_max_pool2d_backward.grad_input(Tensor grad_output, Tensor self, int[2] kernel_size, int[2] output_size, Tensor indices, *, Tensor(a!) grad_input) -> Tensor(a!)
   python_module: nn
+  structured: True
   dispatch:
-    CPU: fractional_max_pool2d_backward_out_cpu
-    CUDA: fractional_max_pool2d_backward_out_cuda
+    CPU: fractional_max_pool2d_backward_cpu
+    CUDA: fractional_max_pool2d_backward_cuda
 
 - func: fractional_max_pool2d_backward(Tensor grad_output, Tensor self, int[2] kernel_size, int[2] output_size, Tensor indices) -> Tensor
   python_module: nn
-  dispatch:
-    CPU: fractional_max_pool2d_backward_cpu
-    CUDA: fractional_max_pool2d_backward_cuda
+  structured_delegate: fractional_max_pool2d_backward.grad_input
 
 # Return: (Tensor output, Tensor indices)
 - func: fractional_max_pool3d.output(Tensor self, int[3] kernel_size, int[3] output_size, Tensor random_samples, *, Tensor(a!) output, Tensor(b!) indices) -> (Tensor(a!), Tensor(b!))
   python_module: nn
+  structured: True
+  precomputed:
+  - kernel_size -> int poolSizeT, int poolSizeH, int poolSizeW
+  - output_size -> int outputT, int outputH, int outputW
   dispatch:
     CPU: fractional_max_pool3d_out_cpu
     CUDA: fractional_max_pool3d_out_cuda
@@ -8953,9 +9289,7 @@
 # Return: (Tensor output, Tensor indices)
 - func: fractional_max_pool3d(Tensor self, int[3] kernel_size, int[3] output_size, Tensor random_samples) -> (Tensor, Tensor)
   python_module: nn
-  dispatch:
-    CPU: fractional_max_pool3d_cpu
-    CUDA: fractional_max_pool3d_cuda
+  structured_delegate: fractional_max_pool3d.output
 
 - func: fractional_max_pool3d_backward.grad_input(Tensor grad_output, Tensor self, int[3] kernel_size, int[3] output_size, Tensor indices, *, Tensor(a!) grad_input) -> Tensor(a!)
   python_module: nn
@@ -9225,6 +9559,16 @@
   dispatch:
     CompositeExplicitAutograd: upsample_bilinear2d_backward
 
+- func: _upsample_bilinear2d_aa.vec(Tensor input, int[]? output_size, bool align_corners, float[]? scale_factors) -> Tensor
+  python_module: nn
+  dispatch:
+    CompositeExplicitAutograd: _upsample_bilinear2d_aa
+
+- func: _upsample_bilinear2d_aa_backward.vec(Tensor grad_output, int[]? output_size, int[] input_size, bool align_corners, float[]? scale_factors) -> Tensor
+  python_module: nn
+  dispatch:
+    CompositeExplicitAutograd: _upsample_bilinear2d_aa_backward
+
 - func: upsample_trilinear3d.vec(Tensor input, int[]? output_size, bool align_corners, float[]? scale_factors) -> Tensor
   python_module: nn
   dispatch:
@@ -9245,26 +9589,56 @@
   dispatch:
     CompositeExplicitAutograd: upsample_bicubic2d_backward
 
+- func: _upsample_bicubic2d_aa.vec(Tensor input, int[]? output_size, bool align_corners, float[]? scale_factors) -> Tensor
+  python_module: nn
+  dispatch:
+    CompositeExplicitAutograd: _upsample_bicubic2d_aa
+
+- func: _upsample_bicubic2d_aa_backward.vec(Tensor grad_output, int[]? output_size, int[] input_size, bool align_corners, float[]? scale_factors) -> Tensor
+  python_module: nn
+  dispatch:
+    CompositeExplicitAutograd: _upsample_bicubic2d_aa_backward
+
 - func: upsample_nearest1d.vec(Tensor input, int[]? output_size, float[]? scale_factors) -> Tensor
   python_module: nn
   dispatch:
     CompositeExplicitAutograd: upsample_nearest1d
 
+- func: _upsample_nearest_exact1d.vec(Tensor input, int[]? output_size, float[]? scale_factors) -> Tensor
+  python_module: nn
+  dispatch:
+    CompositeExplicitAutograd: _upsample_nearest_exact1d
+
 - func: upsample_nearest1d_backward.vec(Tensor grad_output, int[]? output_size, int[] input_size, float[]? scale_factors) -> Tensor
   python_module: nn
   dispatch:
     CompositeExplicitAutograd: upsample_nearest1d_backward
 
+- func: _upsample_nearest_exact1d_backward.vec(Tensor grad_output, int[]? output_size, int[] input_size, float[]? scale_factors) -> Tensor
+  python_module: nn
+  dispatch:
+    CompositeExplicitAutograd: _upsample_nearest_exact1d_backward
+
 - func: upsample_nearest2d.vec(Tensor input, int[]? output_size, float[]? scale_factors) -> Tensor
   python_module: nn
   dispatch:
     CompositeExplicitAutograd: upsample_nearest2d
 
+- func: _upsample_nearest_exact2d.vec(Tensor input, int[]? output_size, float[]? scale_factors) -> Tensor
+  python_module: nn
+  dispatch:
+    CompositeExplicitAutograd: _upsample_nearest_exact2d
+
 - func: upsample_nearest2d_backward.vec(Tensor grad_output, int[]? output_size, int[] input_size, float[]? scale_factors) -> Tensor
   python_module: nn
   dispatch:
     CompositeExplicitAutograd: upsample_nearest2d_backward
 
+- func: _upsample_nearest_exact2d_backward.vec(Tensor grad_output, int[]? output_size, int[] input_size, float[]? scale_factors) -> Tensor
+  python_module: nn
+  dispatch:
+    CompositeExplicitAutograd: _upsample_nearest_exact2d_backward
+
 - func: upsample_nearest3d.vec(Tensor input, int[]? output_size, float[]? scale_factors) -> Tensor
   python_module: nn
   dispatch:
@@ -9272,12 +9646,25 @@
     CUDA: upsample_nearest3d_cuda
     QuantizedCPU: upsample_nearest3d_quantized_cpu
 
+- func: _upsample_nearest_exact3d.vec(Tensor input, int[]? output_size, float[]? scale_factors) -> Tensor
+  python_module: nn
+  dispatch:
+    CPU: _upsample_nearest_exact3d_cpu
+    CUDA: _upsample_nearest_exact3d_cuda
+    QuantizedCPU: _upsample_nearest_exact3d_quantized_cpu
+
 - func: upsample_nearest3d_backward.vec(Tensor grad_output, int[]? output_size, int[] input_size, float[]? scale_factors) -> Tensor
   python_module: nn
   dispatch:
     CPU: upsample_nearest3d_backward_cpu
     CUDA: upsample_nearest3d_backward_cuda
 
+- func: _upsample_nearest_exact3d_backward.vec(Tensor grad_output, int[]? output_size, int[] input_size, float[]? scale_factors) -> Tensor
+  python_module: nn
+  dispatch:
+    CPU: _upsample_nearest_exact3d_backward_cpu
+    CUDA: _upsample_nearest_exact3d_backward_cuda
+
 # NOTE: all of the non-"vec" upsample overloads are only kept for backward compatibility.
 - func: upsample_linear1d.out(Tensor self, int[1] output_size, bool align_corners, float? scales=None, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
@@ -9325,6 +9712,28 @@
   python_module: nn
   structured_delegate: upsample_bilinear2d_backward.grad_input
 
+- func: _upsample_bilinear2d_aa.out(Tensor self, int[2] output_size, bool align_corners, float? scales_h=None, float? scales_w=None, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: nn
+  structured: True
+  dispatch:
+    CPU: _upsample_bilinear2d_aa_out_cpu
+    CUDA: _upsample_bilinear2d_aa_out_cuda
+
+- func: _upsample_bilinear2d_aa(Tensor self, int[2] output_size, bool align_corners, float? scales_h=None, float? scales_w=None) -> Tensor
+  python_module: nn
+  structured_delegate: _upsample_bilinear2d_aa.out
+
+- func: _upsample_bilinear2d_aa_backward.grad_input(Tensor grad_output, int[2] output_size, int[4] input_size, bool align_corners, float? scales_h=None, float? scales_w=None, *, Tensor(a!) grad_input) -> Tensor(a!)
+  python_module: nn
+  structured: True
+  dispatch:
+    CPU: _upsample_bilinear2d_aa_backward_out_cpu
+    CUDA: _upsample_bilinear2d_aa_backward_out_cuda
+
+- func: _upsample_bilinear2d_aa_backward(Tensor grad_output, int[2] output_size, int[4] input_size, bool align_corners, float? scales_h=None, float? scales_w=None) -> Tensor
+  python_module: nn
+  structured_delegate: _upsample_bilinear2d_aa_backward.grad_input
+
 - func: upsample_bicubic2d.out(Tensor self, int[2] output_size, bool align_corners, float? scales_h=None, float? scales_w=None, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
   structured: True
@@ -9347,6 +9756,28 @@
   python_module: nn
   structured_delegate: upsample_bicubic2d_backward.grad_input
 
+- func: _upsample_bicubic2d_aa.out(Tensor self, int[2] output_size, bool align_corners, float? scales_h=None, float? scales_w=None, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: nn
+  structured: True
+  dispatch:
+    CPU: _upsample_bicubic2d_aa_out_cpu
+    CUDA: _upsample_bicubic2d_aa_out_cuda
+
+- func: _upsample_bicubic2d_aa(Tensor self, int[2] output_size, bool align_corners, float? scales_h=None, float? scales_w=None) -> Tensor
+  python_module: nn
+  structured_delegate: _upsample_bicubic2d_aa.out
+
+- func: _upsample_bicubic2d_aa_backward.grad_input(Tensor grad_output, int[2] output_size, int[4] input_size, bool align_corners, float? scales_h=None, float? scales_w=None, *, Tensor(a!) grad_input) -> Tensor(a!)
+  python_module: nn
+  structured: True
+  dispatch:
+    CPU: _upsample_bicubic2d_aa_backward_out_cpu
+    CUDA: _upsample_bicubic2d_aa_backward_out_cuda
+
+- func: _upsample_bicubic2d_aa_backward(Tensor grad_output, int[2] output_size, int[4] input_size, bool align_corners, float? scales_h=None, float? scales_w=None) -> Tensor
+  python_module: nn
+  structured_delegate: _upsample_bicubic2d_aa_backward.grad_input
+
 - func: upsample_trilinear3d.out(Tensor self, int[3] output_size, bool align_corners, float? scales_d=None, float? scales_h=None, float? scales_w=None, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
   structured: True
@@ -9376,10 +9807,21 @@
     CPU: upsample_nearest1d_out_cpu
     CUDA: upsample_nearest1d_out_cuda
 
+- func: _upsample_nearest_exact1d.out(Tensor self, int[1] output_size, float? scales=None, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: nn
+  structured: True
+  dispatch:
+    CPU: _upsample_nearest_exact1d_out_cpu
+    CUDA: _upsample_nearest_exact1d_out_cuda
+
 - func: upsample_nearest1d(Tensor self, int[1] output_size, float? scales=None) -> Tensor
   python_module: nn
   structured_delegate: upsample_nearest1d.out
 
+- func: _upsample_nearest_exact1d(Tensor self, int[1] output_size, float? scales=None) -> Tensor
+  python_module: nn
+  structured_delegate: _upsample_nearest_exact1d.out
+
 - func: upsample_nearest1d_backward.grad_input(Tensor grad_output, int[1] output_size, int[3] input_size, float? scales=None, *, Tensor(a!) grad_input) -> Tensor(a!)
   python_module: nn
   structured: True
@@ -9387,10 +9829,21 @@
     CPU: upsample_nearest1d_backward_out_cpu
     CUDA: upsample_nearest1d_backward_out_cuda
 
+- func: _upsample_nearest_exact1d_backward.grad_input(Tensor grad_output, int[1] output_size, int[3] input_size, float? scales=None, *, Tensor(a!) grad_input) -> Tensor(a!)
+  python_module: nn
+  structured: True
+  dispatch:
+    CPU: _upsample_nearest_exact1d_backward_out_cpu
+    CUDA: _upsample_nearest_exact1d_backward_out_cuda
+
 - func: upsample_nearest1d_backward(Tensor grad_output, int[1] output_size, int[3] input_size, float? scales=None) -> Tensor
   python_module: nn
   structured_delegate: upsample_nearest1d_backward.grad_input
 
+- func: _upsample_nearest_exact1d_backward(Tensor grad_output, int[1] output_size, int[3] input_size, float? scales=None) -> Tensor
+  python_module: nn
+  structured_delegate: _upsample_nearest_exact1d_backward.grad_input
+
 - func: upsample_nearest2d.out(Tensor self, int[2] output_size, float? scales_h=None, float? scales_w=None, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
   structured: True
@@ -9398,12 +9851,25 @@
     CPU: upsample_nearest2d_out_cpu
     CUDA: upsample_nearest2d_out_cuda
 
+- func: _upsample_nearest_exact2d.out(Tensor self, int[2] output_size, float? scales_h=None, float? scales_w=None, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: nn
+  structured: True
+  dispatch:
+    CPU: _upsample_nearest_exact2d_out_cpu
+    CUDA: _upsample_nearest_exact2d_out_cuda
+
 - func: upsample_nearest2d(Tensor self, int[2] output_size, float? scales_h=None, float? scales_w=None) -> Tensor
   python_module: nn
   structured_delegate: upsample_nearest2d.out
   dispatch:
     QuantizedCPU: upsample_nearest2d_quantized_cpu
 
+- func: _upsample_nearest_exact2d(Tensor self, int[2] output_size, float? scales_h=None, float? scales_w=None) -> Tensor
+  python_module: nn
+  structured_delegate: _upsample_nearest_exact2d.out
+  dispatch:
+    QuantizedCPU: _upsample_nearest_exact2d_quantized_cpu
+
 - func: upsample_nearest2d_backward.grad_input(Tensor grad_output, int[2] output_size, int[4] input_size, float? scales_h=None, float? scales_w=None, *, Tensor(a!) grad_input) -> Tensor(a!)
   python_module: nn
   structured: True
@@ -9411,10 +9877,21 @@
     CPU: upsample_nearest2d_backward_out_cpu
     CUDA: upsample_nearest2d_backward_out_cuda
 
+- func: _upsample_nearest_exact2d_backward.grad_input(Tensor grad_output, int[2] output_size, int[4] input_size, float? scales_h=None, float? scales_w=None, *, Tensor(a!) grad_input) -> Tensor(a!)
+  python_module: nn
+  structured: True
+  dispatch:
+    CPU: _upsample_nearest_exact2d_backward_out_cpu
+    CUDA: _upsample_nearest_exact2d_backward_out_cuda
+
 - func: upsample_nearest2d_backward(Tensor grad_output, int[2] output_size, int[4] input_size, float? scales_h=None, float? scales_w=None) -> Tensor
   python_module: nn
   structured_delegate: upsample_nearest2d_backward.grad_input
 
+- func: _upsample_nearest_exact2d_backward(Tensor grad_output, int[2] output_size, int[4] input_size, float? scales_h=None, float? scales_w=None) -> Tensor
+  python_module: nn
+  structured_delegate: _upsample_nearest_exact2d_backward.grad_input
+
 - func: upsample_nearest3d.out(Tensor self, int[3] output_size, float? scales_d=None, float? scales_h=None, float? scales_w=None, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
   structured: True
@@ -9422,12 +9899,25 @@
     CPU: upsample_nearest3d_out_cpu
     CUDA: upsample_nearest3d_out_cuda
 
+- func: _upsample_nearest_exact3d.out(Tensor self, int[3] output_size, float? scales_d=None, float? scales_h=None, float? scales_w=None, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: nn
+  structured: True
+  dispatch:
+    CPU: _upsample_nearest_exact3d_out_cpu
+    CUDA: _upsample_nearest_exact3d_out_cuda
+
 - func: upsample_nearest3d(Tensor self, int[3] output_size, float? scales_d=None, float? scales_h=None, float? scales_w=None) -> Tensor
   python_module: nn
   structured_delegate: upsample_nearest3d.out
   dispatch:
     QuantizedCPU: upsample_nearest3d_quantized_cpu
 
+- func: _upsample_nearest_exact3d(Tensor self, int[3] output_size, float? scales_d=None, float? scales_h=None, float? scales_w=None) -> Tensor
+  python_module: nn
+  structured_delegate: _upsample_nearest_exact3d.out
+  dispatch:
+    QuantizedCPU: _upsample_nearest_exact3d_quantized_cpu
+
 - func: upsample_nearest3d_backward.grad_input(Tensor grad_output, int[3] output_size, int[5] input_size, float? scales_d=None, float? scales_h=None, float? scales_w=None, *, Tensor(a!) grad_input) -> Tensor(a!)
   python_module: nn
   structured: True
@@ -9435,10 +9925,21 @@
     CPU: upsample_nearest3d_backward_out_cpu
     CUDA: upsample_nearest3d_backward_out_cuda
 
+- func: _upsample_nearest_exact3d_backward.grad_input(Tensor grad_output, int[3] output_size, int[5] input_size, float? scales_d=None, float? scales_h=None, float? scales_w=None, *, Tensor(a!) grad_input) -> Tensor(a!)
+  python_module: nn
+  structured: True
+  dispatch:
+    CPU: _upsample_nearest_exact3d_backward_out_cpu
+    CUDA: _upsample_nearest_exact3d_backward_out_cuda
+
 - func: upsample_nearest3d_backward(Tensor grad_output, int[3] output_size, int[5] input_size, float? scales_d=None, float? scales_h=None, float? scales_w=None) -> Tensor
   python_module: nn
   structured_delegate: upsample_nearest3d_backward.grad_input
 
+- func: _upsample_nearest_exact3d_backward(Tensor grad_output, int[3] output_size, int[5] input_size, float? scales_d=None, float? scales_h=None, float? scales_w=None) -> Tensor
+  python_module: nn
+  structured_delegate: _upsample_nearest_exact3d_backward.grad_input
+
 - func: sigmoid_backward.grad_input(Tensor grad_output, Tensor output, *, Tensor(a!) grad_input) -> Tensor(a!)
   python_module: nn
   structured: True
@@ -9501,18 +10002,6 @@
   python_module: nn
   structured_delegate: slow_conv_transpose2d.out
 
-- func: slow_conv_transpose2d_backward.grad_output(Tensor grad_output, Tensor self, Tensor weight, int[2] kernel_size, int[2] stride, int[2] padding, int[2] output_padding, int[2] dilation, Tensor columns, Tensor ones, *, Tensor(a!) grad_input, Tensor(b!) grad_weight, Tensor(c!) grad_bias) -> (Tensor(a!), Tensor(b!), Tensor(c!))
-  python_module: nn
-  dispatch:
-    CPU: slow_conv_transpose2d_backward_out_cpu
-    CUDA: slow_conv_transpose2d_backward_out_cuda
-
-- func: slow_conv_transpose2d_backward.output_mask(Tensor grad_output, Tensor self, Tensor weight, int[2] kernel_size, int[2] stride, int[2] padding, int[2] output_padding, int[2] dilation, Tensor columns, Tensor ones, bool[3] output_mask) -> (Tensor grad_input, Tensor grad_weight, Tensor grad_bias)
-  python_module: nn
-  dispatch:
-    CPU: slow_conv_transpose2d_backward_cpu
-    CUDA: slow_conv_transpose2d_backward_cuda
-
 - func: slow_conv_transpose3d.out(Tensor self, Tensor weight, int[3] kernel_size, Tensor? bias=None, int[3] stride=1, int[3] padding=0, int[3] output_padding=0, int[3] dilation=1, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
   dispatch:
@@ -9525,43 +10014,31 @@
     CPU: slow_conv_transpose3d_cpu
     CUDA: slow_conv_transpose3d_cuda
 
-- func: slow_conv_transpose3d_backward.grad_output(Tensor grad_output, Tensor self, Tensor weight, int[3] kernel_size, int[3] stride, int[3] padding, int[3] output_padding, int[3] dilation, Tensor finput, Tensor fgrad_input, *, Tensor(a!) grad_input, Tensor(b!) grad_weight, Tensor(c!) grad_bias) -> (Tensor(a!), Tensor(b!), Tensor(c!))
-  python_module: nn
-  dispatch:
-    CPU: slow_conv_transpose3d_backward_out_cpu
-    CUDA: slow_conv_transpose3d_backward_out_cuda
-
-- func: slow_conv_transpose3d_backward.output_mask(Tensor grad_output, Tensor self, Tensor weight, int[3] kernel_size, int[3] stride, int[3] padding, int[3] output_padding, int[3] dilation, Tensor finput, Tensor fgrad_input, bool[3] output_mask) -> (Tensor grad_input, Tensor grad_weight, Tensor grad_bias)
-  python_module: nn
-  dispatch:
-    CPU: slow_conv_transpose3d_backward_cpu
-    CUDA: slow_conv_transpose3d_backward_cuda
-
 - func: thnn_conv2d.out(Tensor self, Tensor weight, int[2] kernel_size, Tensor? bias=None, int[2] stride=1, int[2] padding=0, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
 
 - func: thnn_conv2d(Tensor self, Tensor weight, int[2] kernel_size, Tensor? bias=None, int[2] stride=1, int[2] padding=0) -> Tensor
   python_module: nn
 
-- func: thnn_conv2d_forward.output(Tensor self, Tensor weight, int[2] kernel_size, Tensor? bias, int[2] stride, int[2] padding, *, Tensor(a!) output, Tensor(b!) finput, Tensor(c!) fgrad_input) -> (Tensor(a!), Tensor(b!), Tensor(c!))
+- func: _slow_conv2d_forward.output(Tensor self, Tensor weight, int[2] kernel_size, Tensor? bias, int[2] stride, int[2] padding, *, Tensor(a!) output) -> Tensor(a!)
   python_module: nn
   dispatch:
     CPU: slow_conv2d_forward_out_cpu
     CUDA: slow_conv2d_forward_out_cuda
 
-- func: thnn_conv2d_forward(Tensor self, Tensor weight, int[2] kernel_size, Tensor? bias, int[2] stride, int[2] padding) -> (Tensor output, Tensor finput, Tensor fgrad_input)
+- func: _slow_conv2d_forward(Tensor self, Tensor weight, int[2] kernel_size, Tensor? bias, int[2] stride, int[2] padding) -> Tensor
   python_module: nn
   dispatch:
     CPU: slow_conv2d_forward_cpu
     CUDA: slow_conv2d_forward_cuda
 
-- func: thnn_conv2d_backward.grad_input(Tensor grad_output, Tensor self, Tensor weight, int[2] kernel_size, int[2] stride, int[2] padding, Tensor finput, Tensor fgrad_input, *, Tensor(a!) grad_input, Tensor(b!) grad_weight, Tensor(c!) grad_bias) -> (Tensor(a!), Tensor(b!), Tensor(c!))
+- func: _slow_conv2d_backward.grad_input(Tensor grad_output, Tensor self, Tensor weight, int[2] kernel_size, int[2] stride, int[2] padding, *, Tensor(a!) grad_input, Tensor(b!) grad_weight, Tensor(c!) grad_bias) -> (Tensor(a!), Tensor(b!), Tensor(c!))
   python_module: nn
   dispatch:
     CPU: slow_conv2d_backward_out_cpu
     CUDA: slow_conv2d_backward_out_cuda
 
-- func: thnn_conv2d_backward.output_mask(Tensor grad_output, Tensor self, Tensor weight, int[2] kernel_size, int[2] stride, int[2] padding, Tensor finput, Tensor fgrad_input, bool[3] output_mask) -> (Tensor grad_input, Tensor grad_weight, Tensor grad_bias)
+- func: _slow_conv2d_backward.output_mask(Tensor grad_output, Tensor self, Tensor weight, int[2] kernel_size, int[2] stride, int[2] padding, bool[3] output_mask) -> (Tensor grad_input, Tensor grad_weight, Tensor grad_bias)
   python_module: nn
   dispatch:
     CPU: slow_conv2d_backward_cpu
@@ -9578,81 +10055,39 @@
   dispatch:
     CUDA: conv_depthwise2d_cuda
 
-- func: _conv_depthwise2d_backward.grad_input(Tensor grad_output, Tensor self, Tensor weight, int[2] kernel_size, int[2] stride, int[2] padding, int[2] dilation, *, Tensor(a!) grad_input, Tensor(b!) grad_weight) -> (Tensor(a!), Tensor(b!))
-  python_module: nn
-  dispatch:
-    CUDA: conv_depthwise2d_backward_cuda_out
-
-- func: _conv_depthwise2d_backward.output_mask(Tensor grad_output, Tensor self, Tensor weight, int[2] kernel_size, int[2] stride, int[2] padding, int[2] dilation, bool[2] output_mask) -> (Tensor grad_input, Tensor grad_weight)
-  python_module: nn
-  dispatch:
-    CUDA: conv_depthwise2d_backward_cuda
-
 - func: conv_depthwise3d(Tensor self, Tensor weight, int[3] kernel_size, Tensor? bias, int[3] stride, int[3] padding, int[3] dilation) -> Tensor
   python_module: nn
   dispatch:
     CUDA: conv_depthwise3d_cuda
 
-- func: conv_depthwise3d_backward.grad_input(Tensor grad_output, Tensor self, Tensor weight, int[3] kernel_size, int[3] stride, int[3] padding, int[3] dilation, *, Tensor(a!) grad_input, Tensor(b!) grad_weight, Tensor(c!) grad_bias) -> (Tensor(a!), Tensor(b!), Tensor(c!))
-  python_module: nn
-  dispatch:
-    CUDA: conv_depthwise3d_backward_cuda_out
-
-- func: conv_depthwise3d_backward.output_mask(Tensor grad_output, Tensor self, Tensor weight, int[3] kernel_size, int[3] stride, int[3] padding, int[3] dilation, bool[3] output_mask) -> (Tensor grad_input, Tensor grad_weight, Tensor grad_bias)
-  python_module: nn
-  dispatch:
-    CUDA: conv_depthwise3d_backward_cuda
-
 - func: slow_conv3d.out(Tensor self, Tensor weight, int[3] kernel_size, Tensor? bias=None, int[3] stride=1, int[3] padding=0, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
 
 - func: slow_conv3d(Tensor self, Tensor weight, int[3] kernel_size, Tensor? bias=None, int[3] stride=1, int[3] padding=0) -> Tensor
   python_module: nn
 
-- func: slow_conv3d_forward.output(Tensor self, Tensor weight, int[3] kernel_size, Tensor? bias, int[3] stride, int[3] padding, *, Tensor(a!) output, Tensor(b!) finput, Tensor(c!) fgrad_input) -> (Tensor(a!), Tensor(b!), Tensor(c!))
+- func: slow_conv3d_forward.output(Tensor self, Tensor weight, int[3] kernel_size, Tensor? bias, int[3] stride, int[3] padding, *, Tensor(a!) output) -> Tensor(a!)
   python_module: nn
   dispatch:
     CPU: slow_conv3d_forward_out_cpu
 
-- func: slow_conv3d_forward(Tensor self, Tensor weight, int[3] kernel_size, Tensor? bias, int[3] stride, int[3] padding) -> (Tensor output, Tensor finput, Tensor fgrad_input)
+- func: slow_conv3d_forward(Tensor self, Tensor weight, int[3] kernel_size, Tensor? bias, int[3] stride, int[3] padding) -> Tensor
   python_module: nn
   dispatch:
     CPU: slow_conv3d_forward_cpu
 
-- func: slow_conv3d_backward.grad_input(Tensor grad_output, Tensor self, Tensor weight, int[3] kernel_size, int[3] stride, int[3] padding, Tensor finput, Tensor fgrad_input, *, Tensor(a!) grad_input, Tensor(b!) grad_weight, Tensor(c!) grad_bias) -> (Tensor(a!), Tensor(b!), Tensor(c!))
-  python_module: nn
-  dispatch:
-    CPU: slow_conv3d_backward_out_cpu
-
-- func: slow_conv3d_backward.output_mask(Tensor grad_output, Tensor self, Tensor weight, int[3] kernel_size, int[3] stride, int[3] padding, Tensor finput, Tensor fgrad_input, bool[3] output_mask) -> (Tensor grad_input, Tensor grad_weight, Tensor grad_bias)
-  python_module: nn
-  dispatch:
-    CPU: slow_conv3d_backward_cpu
-
 - func: slow_conv_dilated2d(Tensor self, Tensor weight, int[2] kernel_size, Tensor? bias=None, int[2] stride=1, int[2] padding=0, int[2] dilation=1) -> Tensor
   python_module: nn
   dispatch:
     CPU: slow_conv_dilated2d_cpu
     CUDA: slow_conv_dilated2d_cuda
 
-- func: slow_conv_dilated2d_backward(Tensor grad_output, Tensor self, Tensor weight, int[2] kernel_size, int[2] stride, int[2] padding, int[2] dilation, bool[3] output_mask) -> (Tensor grad_input, Tensor grad_weight, Tensor grad_bias)
-  python_module: nn
-  dispatch:
-    CPU: slow_conv_dilated2d_backward_cpu
-    CUDA: slow_conv_dilated2d_backward_cuda
-
 - func: slow_conv_dilated3d(Tensor self, Tensor weight, int[3] kernel_size, Tensor? bias=None, int[3] stride=1, int[3] padding=0, int[3] dilation=1) -> Tensor
   python_module: nn
   dispatch:
     CPU: slow_conv_dilated3d_cpu
     CUDA: slow_conv_dilated3d_cuda
 
-- func: slow_conv_dilated3d_backward(Tensor grad_output, Tensor self, Tensor weight, int[3] kernel_size, int[3] stride, int[3] padding, int[3] dilation, bool[3] output_mask) -> (Tensor grad_input, Tensor grad_weight, Tensor grad_bias)
-  python_module: nn
-  dispatch:
-    CPU: slow_conv_dilated3d_backward_cpu
-    CUDA: slow_conv_dilated3d_backward_cuda
-
 - func: col2im.out(Tensor self, int[2] output_size, int[2] kernel_size, int[2] dilation, int[2] padding, int[2] stride, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
   dispatch:
@@ -9714,6 +10149,10 @@
   variants: function, method
   device_check: NoCheck
   device_guard: False
+  dispatch:
+    CompositeExplicitAutograd: isinf
+    SparseCPU, SparseCUDA: isinf_sparse
+    SparseCsrCPU, SparseCsrCUDA: isinf_sparse_csr
 
 - func: record_stream(Tensor(a!) self, Stream s) -> ()
   variants: method
@@ -9723,22 +10162,32 @@
 - func: isposinf(Tensor self) -> Tensor
   variants: function, method
   structured_delegate: isposinf.out
+  dispatch:
+    SparseCPU, SparseCUDA: isposinf_sparse
+    SparseCsrCPU, SparseCsrCUDA: isposinf_sparse_csr
 
 - func: isposinf.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: isposinf_out
+    SparseCPU, SparseCUDA: isposinf_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: isposinf_sparse_csr_out
 
 - func: isneginf(Tensor self) -> Tensor
   variants: function, method
   structured_delegate: isneginf.out
+  dispatch:
+    SparseCPU, SparseCUDA: isneginf_sparse
+    SparseCsrCPU, SparseCsrCUDA: isneginf_sparse_csr
 
 - func: isneginf.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: isneginf_out
+    SparseCPU, SparseCUDA: isneginf_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: isneginf_sparse_csr_out
 
 # NOTE [_add_batch_dim and _remove_batch_dim]
 # _add_batch_dim and _remove_batch_dim are meant to be used in the implementation
@@ -10065,11 +10514,11 @@
   python_module: special
   variants: function
 
-- func: special_round(Tensor self) -> Tensor
+- func: special_round(Tensor self, *, int decimals=0) -> Tensor
   python_module: special
   variants: function
 
-- func: special_round.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+- func: special_round.out(Tensor self, *, int decimals=0, Tensor(a!) out) -> Tensor(a!)
   python_module: special
   variants: function
 
@@ -10109,6 +10558,10 @@
   python_module: special
   variants: function
 
+- func: special_softmax(Tensor self, int dim, ScalarType? dtype=None) -> Tensor
+  python_module: special
+  variants: function
+
 ## Functions related to the fast Fourier transform and the torch.fft namespace
 # Note [FFT namespace binding]
 # Functions in the fft python module should have their names start with
@@ -10200,6 +10653,26 @@
   python_module: fft
   variants: function
 
+- func: fft_hfft2(Tensor self, int[1]? s=None, int[1] dim=[-2,-1], str? norm=None) -> Tensor
+  use_const_ref_for_mutable_tensors: True
+  python_module: fft
+  variants: function
+
+- func: fft_hfft2.out(Tensor self, int[1]? s=None, int[1] dim=[-2,-1], str? norm=None, *, Tensor(a!) out) -> Tensor(a!)
+  use_const_ref_for_mutable_tensors: True
+  python_module: fft
+  variants: function
+
+- func: fft_ihfft2(Tensor self, int[1]? s=None, int[1] dim=[-2,-1], str? norm=None) -> Tensor
+  use_const_ref_for_mutable_tensors: True
+  python_module: fft
+  variants: function
+
+- func: fft_ihfft2.out(Tensor self, int[1]? s=None, int[1] dim=[-2,-1], str? norm=None, *, Tensor(a!) out) -> Tensor(a!)
+  use_const_ref_for_mutable_tensors: True
+  python_module: fft
+  variants: function
+
 - func: fft_fftn(Tensor self, int[1]? s=None, int[1]? dim=None, str? norm=None) -> Tensor
   python_module: fft
   variants: function
@@ -10232,6 +10705,26 @@
   python_module: fft
   variants: function
 
+- func: fft_hfftn(Tensor self, int[1]? s=None, int[1]? dim=None, str? norm=None) -> Tensor
+  use_const_ref_for_mutable_tensors: True
+  python_module: fft
+  variants: function
+
+- func: fft_hfftn.out(Tensor self, int[1]? s=None, int[1]? dim=None, str? norm=None, *, Tensor(a!) out) -> Tensor(a!)
+  use_const_ref_for_mutable_tensors: True
+  python_module: fft
+  variants: function
+
+- func: fft_ihfftn(Tensor self, int[1]? s=None, int[1]? dim=None, str? norm=None) -> Tensor
+  use_const_ref_for_mutable_tensors: True
+  python_module: fft
+  variants: function
+
+- func: fft_ihfftn.out(Tensor self, int[1]? s=None, int[1]? dim=None, str? norm=None, *, Tensor(a!) out) -> Tensor(a!)
+  use_const_ref_for_mutable_tensors: True
+  python_module: fft
+  variants: function
+
 - func: fft_fftfreq(int n, float d=1.0, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
   python_module: fft
   variants: function
@@ -10286,6 +10779,38 @@
   python_module: linalg
   variants: function
 
+- func: linalg_cross(Tensor self, Tensor other, *, int dim=-1) -> Tensor
+  python_module: linalg
+  variants: function
+  dispatch:
+    CPU, CUDA: linalg_cross
+
+- func: linalg_cross.out(Tensor self, Tensor other, *, int dim=-1, Tensor(a!) out) -> Tensor(a!)
+  python_module: linalg
+  dispatch:
+    CPU, CUDA: linalg_cross_out
+
+# linalg.lu_factor
+- func: linalg_lu_factor(Tensor A, *, bool pivot=True) -> (Tensor LU, Tensor pivots)
+  python_module: linalg
+  variants: function
+
+- func: linalg_lu_factor.out(Tensor A, *, bool pivot=True, Tensor(a!) LU, Tensor(b!) pivots) -> (Tensor(a!) LU, Tensor(b!) pivots)
+  python_module: linalg
+  variants: function
+
+- func: linalg_lu_factor_ex(Tensor A, *, bool pivot=True, bool check_errors=False) -> (Tensor LU, Tensor pivots, Tensor info)
+  python_module: linalg
+  structured_delegate: linalg_lu_factor_ex.out
+  variants: function
+
+- func: linalg_lu_factor_ex.out(Tensor A, *, bool pivot=True, bool check_errors=False, Tensor(a!) LU, Tensor(b!) pivots, Tensor(c!) info) -> (Tensor(a!) LU, Tensor(b!) pivots, Tensor(c!) info)
+  python_module: linalg
+  variants: function
+  structured: True
+  dispatch:
+    CPU, CUDA: linalg_lu_factor_ex_out
+
 - func: linalg_det(Tensor self) -> Tensor
   python_module: linalg
   variants: function
@@ -10327,6 +10852,12 @@
 - func: linalg_matmul.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   python_module: linalg
 
+- func: linalg_matrix_exp(Tensor self) -> Tensor
+  python_module: linalg
+  variants: function
+  dispatch:
+    CPU, CUDA: linalg_matrix_exp
+
 - func: linalg_slogdet(Tensor self) -> (Tensor sign, Tensor logabsdet)
   python_module: linalg
   variants: function
@@ -10467,18 +10998,30 @@
 - func: linalg_matrix_norm.str_ord_out(Tensor self, str ord='fro', int[] dim=[-2,-1], bool keepdim=False, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)
   python_module: linalg
 
-- func: linalg_svd.U(Tensor self, bool full_matrices=True, *, Tensor(a!) U, Tensor(b!) S, Tensor(c!) Vh) -> (Tensor(a!) U, Tensor(b!) S, Tensor(c!) Vh)
+# This function is exposes the `compute_uv` flag, which is then used to implement `linalg.svd` and
+# `linalg.svdvals` as composite functions that call this one
+- func: _linalg_svd(Tensor A, bool full_matrices=False, bool compute_uv=True) -> (Tensor U, Tensor S, Tensor Vh)
+  variants: function
+  structured_delegate: _linalg_svd.U
+
+- func: _linalg_svd.U(Tensor A, bool full_matrices=False, bool compute_uv=True, *, Tensor(a!) U, Tensor(b!) S, Tensor(c!) Vh) -> (Tensor(a!) U, Tensor(b!) S, Tensor(c!) Vh)
+  structured: True
+  dispatch:
+    CPU, CUDA: _linalg_svd_out
+
+- func: linalg_svd(Tensor A, bool full_matrices=True) -> (Tensor U, Tensor S, Tensor Vh)
   python_module: linalg
+  variants: function
 
-- func: linalg_svd(Tensor self, bool full_matrices=True) -> (Tensor U, Tensor S, Tensor Vh)
+- func: linalg_svd.U(Tensor A, bool full_matrices=True, *, Tensor(a!) U, Tensor(b!) S, Tensor(c!) Vh) -> (Tensor(a!) U, Tensor(b!) S, Tensor(c!) Vh)
   python_module: linalg
   variants: function
 
-- func: linalg_svdvals(Tensor input) -> Tensor
+- func: linalg_svdvals(Tensor A) -> Tensor
   python_module: linalg
   variants: function
 
-- func: linalg_svdvals.out(Tensor input, *, Tensor(a!) out) -> Tensor(a!)
+- func: linalg_svdvals.out(Tensor A, *, Tensor(a!) out) -> Tensor(a!)
   python_module: linalg
   variants: function
 
@@ -10498,7 +11041,29 @@
   python_module: linalg
   variants: function
 
-- func: linalg_pinv(Tensor self, float rcond=1e-15, bool hermitian=False) -> Tensor
+- func: linalg_pinv.atol_rtol_tensor(Tensor self, *, Tensor? atol=None, Tensor? rtol=None, bool hermitian=False) -> Tensor
+  python_module: linalg
+  variants: function
+  dispatch:
+    CompositeExplicitAutograd: linalg_pinv
+
+- func: linalg_pinv.atol_rtol_tensor_out(Tensor self, *, Tensor? atol=None, Tensor? rtol=None, bool hermitian=False, Tensor(a!) out) -> Tensor(a!)
+  python_module: linalg
+  variants: function
+  dispatch:
+    CompositeExplicitAutograd: linalg_pinv_out
+
+- func: linalg_pinv.atol_rtol_float(Tensor self, *, float? atol=None, float? rtol=None, bool hermitian=False) -> Tensor
+  cpp_no_default_args: ['atol', 'rtol']
+  python_module: linalg
+  variants: function
+
+- func: linalg_pinv.atol_rtol_float_out(Tensor self, *, float? atol=None, float? rtol=None, bool hermitian=False, Tensor(a!) out) -> Tensor(a!)
+  cpp_no_default_args: ['atol', 'rtol']
+  python_module: linalg
+  variants: function
+
+- func: linalg_pinv(Tensor self, float rcond, bool hermitian=False) -> Tensor
   python_module: linalg
   variants: function
 
@@ -10506,7 +11071,7 @@
   python_module: linalg
   variants: function
 
-- func: linalg_pinv.out(Tensor self, float rcond=1e-15, bool hermitian=False, *, Tensor(a!) out) -> Tensor(a!)
+- func: linalg_pinv.out(Tensor self, float rcond, bool hermitian=False, *, Tensor(a!) out) -> Tensor(a!)
   python_module: linalg
   variants: function
 
@@ -10565,11 +11130,29 @@
 - func: linalg_matrix_power.out(Tensor self, int n, *, Tensor(a!) out) -> Tensor(a!)
   python_module: linalg
 
-- func: linalg_matrix_rank(Tensor self, float? tol=None, bool hermitian=False) -> Tensor
+- func: linalg_matrix_rank.atol_rtol_tensor(Tensor input, *, Tensor? atol=None, Tensor? rtol=None, bool hermitian=False) -> Tensor
   python_module: linalg
   variants: function
 
-- func: linalg_matrix_rank.out(Tensor self, float? tol=None, bool hermitian=False, *, Tensor(a!) out) -> Tensor(a!)
+- func: linalg_matrix_rank.atol_rtol_tensor_out(Tensor input, *, Tensor? atol=None, Tensor? rtol=None, bool hermitian=False, Tensor(a!) out) -> Tensor(a!)
+  python_module: linalg
+  variants: function
+
+- func: linalg_matrix_rank.atol_rtol_float(Tensor self, *, float? atol=None, float? rtol=None, bool hermitian=False) -> Tensor
+  cpp_no_default_args: ['atol', 'rtol']
+  python_module: linalg
+  variants: function
+
+- func: linalg_matrix_rank.atol_rtol_float_out(Tensor self, *, float? atol=None, float? rtol=None, bool hermitian=False, Tensor(a!) out) -> Tensor(a!)
+  cpp_no_default_args: ['atol', 'rtol']
+  python_module: linalg
+  variants: function
+
+- func: linalg_matrix_rank(Tensor self, float tol, bool hermitian=False) -> Tensor
+  python_module: linalg
+  variants: function
+
+- func: linalg_matrix_rank.out(Tensor self, float tol, bool hermitian=False, *, Tensor(a!) out) -> Tensor(a!)
   python_module: linalg
   variants: function
 
@@ -10622,6 +11205,12 @@
   cpp_no_default_args: ['a', 'b']
   python_module: nn
 
+# Note: this function is only for testing.
+- func: _test_warn_in_autograd(Tensor self) -> Tensor
+  python_module: nn
+  dispatch:
+    CompositeExplicitAutograd: _test_warn_in_autograd
+
 - func: segment_reduce(Tensor data, str reduce, *, Tensor? lengths=None, Tensor? indices=None, int axis=0, bool unsafe=False, Scalar? initial=None) -> Tensor
   variants: function
   dispatch: